Egothor/Radixor
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Refresh multilingual benchmarks and fix overlapping gold evaluation

Browse Source 
Recompute published benchmark results for all default language models,
exclude Polish Polimorf, add Hebrew documentation, and record the current
benchmark environment. Evaluate repeated surface forms as an overlapping
gold cover and publish only applicable metrics for candidate policies.
This commit is contained in:
Leo Galambos
2026-07-23 17:06:41 +02:00
parent 1f1b03c6a8
commit b29699b763
64 changed files with 5392 additions and 4765 deletions
Download Patch File Download Diff File Expand all files Collapse all files

319
docs/benchmarks/languages/czech.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `CS_CZ` | 5,113 | 56,612 | 10,049 | 46,563 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `cs-cz-default` | `1.0.0` | `CS_CZ` | 5,113 | 56,612 | 10,049 | 46,563 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **56,612**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **56,612**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 675 | 1.192% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 22,681 | 40.064% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 14,980 | 26.461% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 10,109 | 17.857% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 8,167 | 14.426% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 711 | 1.256% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 22,643 | 39.997% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 15,007 | 26.509% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 10,046 | 17.745% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 8,205 | 14.493% |
## Accuracy
@@ -34,15 +34,25 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene HunspellStemFilter | 84.850% | 82.269% | 96.806% | Benchmark-only Czech Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene CzechStemFilter | 16.784% | 15.538% | 22.559% | Lucene Czech suffix stemmer implemented as a TokenFilter. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `czechRadixor` | 3.332 | 0.240 | 71.6 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 346.819 | 3.622 | 7448.4 | 104.091 | Benchmark-only Czech Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene CzechStemFilter | `czechLuceneCzechStemFilter` | 3.163 | 0.253 | 67.9 | 0.949 | Czech suffix stemmer implemented as a Lucene TokenFilter. |
| Radixor | `czechRadixor` | 3.395 | 0.066 | 72.9 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 381.189 | 32.563 | 8186.5 | 112.265 | Benchmark-only Czech Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene CzechStemFilter | `czechLuceneCzechStemFilter` | 3.125 | 0.042 | 67.1 | 0.920 | Czech suffix stemmer implemented as a Lucene TokenFilter. |
## Interpretation Notes
@@ -56,29 +66,29 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `CS_CZ` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `CS_CZ` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/cs_cz/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `cs-cz-default`, loaded from classpath resource `org/egothor/stemmer/models/cs-cz-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.996565** among 3 deterministic stemmers. The runner-up is `HUNSPELL CZECH LUCENE FILTER` at 0.853752, a difference of 0.142813. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.997139** among 3 deterministic stemmers. The runner-up is `HUNSPELL CZECH LUCENE FILTER` at 0.852770, a difference of 0.144369. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.996617** among 3 deterministic stemmers. The runner-up is `HUNSPELL CZECH LUCENE FILTER` at 0.854132, a difference of 0.142485. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.997195** among 3 deterministic stemmers. The runner-up is `HUNSPELL CZECH LUCENE FILTER` at 0.853150, a difference of 0.144045. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996565|3867 / 1334876815 (0.000290%)|2073 / 301835 (0.686799%)|0.988432|0.990189|0.990191|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.853752|11408 / 1334876815 (0.000855%)|88283 / 301835 (29.248762%)|0.888560|0.810759|0.819499|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.793614|14480 / 1334876815 (0.001085%)|124586 / 301835 (41.276194%)|0.829234|0.718241|0.736765|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.996617|0.000000%|0.676519%|
|2|HUNSPELL CZECH LUCENE FILTER|0.854132|0.000691%|29.172837%|
|3|CZECH LUCENE CZECH STEM FILTER|0.794343|0.000928%|41.130549%|
</div>
@@ -86,9 +96,9 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987264|0.993132|0.999997|0.996565|0.999996|0.000004|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.949289|0.707512|0.999991|0.853752|0.999925|0.000075|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.924477|0.587238|0.999989|0.793614|0.999896|0.000104|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.993235|1.000000|0.996617|0.999998|0.000002|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.958877|0.708272|0.999993|0.854132|0.999927|0.000073|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.935210|0.588695|0.999991|0.794343|0.999897|0.000103|
</details>
@@ -96,19 +106,9 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988432|0.990189|0.991953|0.980569|0.990194|0.990191|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.888560|0.810759|0.745486|0.681745|0.819533|0.819499|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.829234|0.718241|0.633453|0.560356|0.736809|0.736765|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.990187|0.998733|0.998686|0.998709|0.998709|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.810723|0.995777|0.952852|0.973842|0.973842|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.718192|0.993801|0.944977|0.968774|0.968774|
|1|Radixor|PRIMARY_OUTPUT|0.998640|0.996606|0.994581|0.993235|0.996612|0.996611|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.895506|0.814739|0.747335|0.687392|0.824103|0.824070|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.836710|0.722556|0.635811|0.565626|0.741992|0.741949|
</details>
@@ -116,67 +116,42 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|299762|3867|2073|1334872948|3867 / 1334876815|2073 / 301835|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|213552|11408|88283|1334865407|11408 / 1334876815|88283 / 301835|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|177249|14480|124586|1334862335|14480 / 1334876815|124586 / 301835|
|1|Radixor|PRIMARY_OUTPUT|298476|0|2033|1320705191|0 / 1320705191|2033 / 300509|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|212842|9128|87667|1320696063|9128 / 1320705191|87667 / 300509|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|176908|12256|123601|1320692935|12256 / 1320705191|123601 / 300509|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 1334876815 (0.000000%)|0 / 301835 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|0.871577|10102 / 1334876815 (0.000757%)|77523 / 301835 (25.683900%)|0.904855|0.836596|0.843258|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
|HUNSPELL CZECH LUCENE FILTER|0.000650%|25.611213%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|0.956905|0.743161|0.999992|0.871577|0.999934|0.000066|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|0.904855|0.836596|0.777914|0.719094|0.843288|0.843258|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|301835|0|0|1334876815|0 / 1334876815|0 / 301835|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|224312|10102|77523|1334866713|10102 / 1334876815|77523 / 301835|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 1320705191|0 / 300509|
|HUNSPELL CZECH LUCENE FILTER|8582 / 1320705191|76964 / 300509|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999998|5850 / 1334876815 (0.000438%)|0 / 301835 (0.000000%)|0.984732|0.990402|0.990446|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.871575|13917 / 1334876815 (0.001043%)|77523 / 301835 (25.683900%)|0.893851|0.830687|0.836477|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|HUNSPELL CZECH LUCENE FILTER|0.871940|0.000816%|25.611213%|
</div>
@@ -184,8 +159,8 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.980987|1.000000|0.999996|0.999998|0.999996|0.000004|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.941581|0.743161|0.999990|0.871575|0.999932|0.000068|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.954016|0.743888|0.999992|0.871940|0.999934|0.000066|
</details>
@@ -193,17 +168,8 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.984732|0.990402|0.996139|0.980987|0.990448|0.990446|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.893851|0.830687|0.775861|0.710406|0.836509|0.836477|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.903001|0.835949|0.778167|0.718138|0.842426|0.842395|
</details>
@@ -211,8 +177,8 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|301835|5850|0|1334870965|5850 / 1334876815|0 / 301835|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|224312|13917|77523|1334862898|13917 / 1334876815|77523 / 301835|
|1|Radixor|ALL_CANDIDATES|300509|0|0|1320705191|0 / 1320705191|0 / 300509|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|223545|10775|76964|1320694416|10775 / 1320705191|76964 / 300509|
</details>
@@ -222,22 +188,22 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|2073|3867|1983|596|1.153340%|4|52319|
|HUNSPELL CZECH LUCENE FILTER|10760|1306|2509|3317|6.418840%|5|55596|
|Radixor|2033|0|0|321|0.624501%|4|51739|
|HUNSPELL CZECH LUCENE FILTER|10703|546|1647|3194|6.213887%|5|55179|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.997139|3863 / 1298544215 (0.000297%)|1709 / 298813 (0.571930%)|0.988580|0.990710|0.990714|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.852770|11239 / 1298544215 (0.000866%)|87986 / 298813 (29.445171%)|0.888009|0.809505|0.818403|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.791794|13950 / 1298544215 (0.001074%)|124426 / 298813 (41.640089%)|0.828709|0.715948|0.735055|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.997195|0.000000%|0.561033%|
|2|HUNSPELL CZECH LUCENE FILTER|0.853150|0.000700%|29.369351%|
|3|CZECH LUCENE CZECH STEM FILTER|0.792522|0.000918%|41.494586%|
</div>
@@ -245,9 +211,9 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987165|0.994281|0.999997|0.997139|0.999996|0.000004|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.949389|0.705548|0.999991|0.852770|0.999924|0.000076|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.925931|0.583599|0.999989|0.791794|0.999893|0.000107|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.994390|1.000000|0.997195|0.999999|0.000001|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.958957|0.706306|0.999993|0.853150|0.999925|0.000075|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.936557|0.585054|0.999991|0.792522|0.999895|0.000105|
</details>
@@ -255,19 +221,9 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988580|0.990710|0.992849|0.981591|0.990716|0.990714|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.888009|0.809505|0.743753|0.679973|0.818437|0.818403|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.828709|0.715948|0.630198|0.557569|0.735100|0.735055|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.990708|0.998726|0.999030|0.998878|0.998878|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.809467|0.995812|0.952394|0.973619|0.973619|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.715897|0.993897|0.944297|0.968463|0.968463|
|1|Radixor|PRIMARY_OUTPUT|0.998873|0.997187|0.995507|0.994390|0.997191|0.997190|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|0.894932|0.813466|0.745594|0.685581|0.822993|0.822960|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|0.836092|0.720206|0.632534|0.562751|0.740227|0.740184|
</details>
@@ -275,67 +231,42 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|297104|3863|1709|1298540352|3863 / 1298544215|1709 / 298813|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|210827|11239|87986|1298532976|11239 / 1298544215|87986 / 298813|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|174387|13950|124426|1298530265|13950 / 1298544215|124426 / 298813|
|1|Radixor|PRIMARY_OUTPUT|295818|0|1669|1284770069|0 / 1284770069|1669 / 297487|
|2|HUNSPELL CZECH LUCENE FILTER|PRIMARY_OUTPUT|210117|8993|87370|1284761076|8993 / 1284770069|87370 / 297487|
|3|CZECH LUCENE CZECH STEM FILTER|PRIMARY_OUTPUT|174046|11790|123441|1284758279|11790 / 1284770069|123441 / 297487|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 1298544215 (0.000000%)|0 / 298813 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|0.870432|10028 / 1298544215 (0.000772%)|77431 / 298813 (25.912862%)|0.904004|0.835052|0.841852|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
|HUNSPELL CZECH LUCENE FILTER|0.000663%|25.840457%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|0.956666|0.740871|0.999992|0.870432|0.999933|0.000067|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|0.904004|0.835052|0.775874|0.716815|0.841883|0.841852|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|298813|0|0|1298544215|0 / 1298544215|0 / 298813|
|2|HUNSPELL CZECH LUCENE FILTER|ANY_CANDIDATE|221382|10028|77431|1298534187|10028 / 1298544215|77431 / 298813|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 1284770069|0 / 297487|
|HUNSPELL CZECH LUCENE FILTER|8518 / 1284770069|76872 / 297487|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999998|5782 / 1298544215 (0.000445%)|0 / 298813 (0.000000%)|0.984756|0.990418|0.990461|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.870430|13601 / 1298544215 (0.001047%)|77431 / 298813 (25.912862%)|0.893574|0.829463|0.835425|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|HUNSPELL CZECH LUCENE FILTER|0.870794|0.000819%|25.840457%|
</div>
@@ -343,8 +274,8 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.981017|1.000000|0.999996|0.999998|0.999996|0.000004|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.942119|0.740871|0.999990|0.870430|0.999930|0.000070|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.954473|0.741595|0.999992|0.870794|0.999932|0.000068|
</details>
@@ -352,17 +283,8 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.984756|0.990418|0.996145|0.981017|0.990463|0.990461|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.893574|0.829463|0.773936|0.708617|0.835457|0.835425|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|0.902651|0.834675|0.776220|0.716259|0.841328|0.841297|
</details>
@@ -370,8 +292,8 @@ This mode contains **7 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|298813|5782|0|1298538433|5782 / 1298544215|0 / 298813|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|221382|13601|77431|1298530614|13601 / 1298544215|77431 / 298813|
|1|Radixor|ALL_CANDIDATES|297487|0|0|1284770069|0 / 1284770069|0 / 297487|
|2|HUNSPELL CZECH LUCENE FILTER|ALL_CANDIDATES|220615|10523|76872|1284759546|10523 / 1284770069|76872 / 297487|
</details>
@@ -381,20 +303,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|1709|3863|1919|540|1.059488%|4|51543|
|HUNSPELL CZECH LUCENE FILTER|10555|1211|2362|3237|6.351044%|5|54804|
|Radixor|1669|0|0|269|0.530603%|4|50975|
|HUNSPELL CZECH LUCENE FILTER|10498|475|1530|3117|6.148293%|5|54394|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -403,16 +325,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `CS_CZ`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

279
docs/benchmarks/languages/danish.md
View File

@@ -8,20 +8,20 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `DA_DK` | 4,179 | 32,256 | 8,356 | 23,900 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `da-dk-default` | `1.0.0` | `DA_DK` | 4,179 | 32,256 | 8,356 | 23,900 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **32,256**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **32,256**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 137 | 0.425% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 179 | 0.555% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 1,127 | 3.494% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 22,586 | 70.021% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 8,405 | 26.057% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 22,680 | 70.312% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 8,269 | 25.636% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 1 | 0.003% |
## Accuracy
@@ -34,15 +34,25 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene SnowballFilter | 55.509% | 54.159% | 59.371% | Lucene TokenFilter integration path around the Snowball algorithm. |
| Official Snowball direct | 55.509% | 54.159% | 59.371% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `radixor[DANISH]` | 1.143 | 0.017 | 47.8 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Official Snowball direct | `snowballDirect[DANISH]` | 2.168 | 0.058 | 90.7 | 1.896 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[DANISH]` | 2.975 | 0.143 | 124.5 | 2.602 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `radixor[DANISH]` | 1.206 | 0.134 | 50.5 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Official Snowball direct | `snowballDirect[DANISH]` | 2.326 | 0.205 | 97.3 | 1.928 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[DANISH]` | 3.275 | 0.335 | 137.0 | 2.716 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -56,29 +66,29 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `DA_DK` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `DA_DK` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/da_dk/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `da-dk-default`, loaded from classpath resource `org/egothor/stemmer/models/da-dk-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.996066** among 3 deterministic stemmers. The runner-up is `SNOWBALL DANISH LUCENE FILTER` at 0.937969, a difference of 0.058097. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.996305** among 3 deterministic stemmers. The runner-up is `SNOWBALL DANISH DIRECT` at 0.938074, a difference of 0.058230. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.996243** among 3 deterministic stemmers. The runner-up is `SNOWBALL DANISH LUCENE FILTER` at 0.937905, a difference of 0.058337. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.996482** among 3 deterministic stemmers. The runner-up is `SNOWBALL DANISH DIRECT` at 0.938010, a difference of 0.058472. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996066|1165 / 394111186 (0.000296%)|707 / 89895 (0.786473%)|0.988108|0.989614|0.989615|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.937969|6507 / 394111186 (0.001651%)|11151 / 89895 (12.404472%)|0.913718|0.899181|0.899475|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.937903|6341 / 394111186 (0.001609%)|11163 / 89895 (12.417821%)|0.915090|0.899959|0.900279|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.996243|0.000000%|0.751435%|
|2|SNOWBALL DANISH LUCENE FILTER|0.937905|0.001273%|12.417638%|
|3|SNOWBALL DANISH DIRECT|0.937839|0.001230%|12.431016%|
</div>
@@ -86,9 +96,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987106|0.992135|0.999997|0.996066|0.999995|0.000005|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.923672|0.875955|0.999983|0.937969|0.999955|0.000045|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.925464|0.875822|0.999984|0.937903|0.999956|0.000044|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.992486|1.000000|0.996243|0.999998|0.000002|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.940600|0.875824|0.999987|0.937905|0.999959|0.000041|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.942465|0.875690|0.999988|0.937839|0.999959|0.000041|
</details>
@@ -96,19 +106,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988108|0.989614|0.991125|0.979442|0.989618|0.989615|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.913718|0.899181|0.885100|0.816830|0.899498|0.899475|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.915090|0.899959|0.885320|0.818114|0.900301|0.900279|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989612|0.998466|0.998719|0.998592|0.998592|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.899159|0.994053|0.978603|0.986268|0.986268|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.899937|0.994196|0.978579|0.986326|0.986326|
|1|Radixor|PRIMARY_OUTPUT|0.998488|0.996229|0.993979|0.992486|0.996236|0.996235|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.926889|0.907057|0.888055|0.829921|0.907634|0.907614|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.928307|0.907851|0.888277|0.831252|0.908464|0.908444|
</details>
@@ -116,61 +116,39 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|89188|1165|707|394110021|1165 / 394111186|707 / 89895|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|78744|6507|11151|394104679|6507 / 394111186|11151 / 89895|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|78732|6341|11163|394104845|6341 / 394111186|11163 / 89895|
|1|Radixor|PRIMARY_OUTPUT|89021|0|674|389687465|0 / 389687465|674 / 89695|
|2|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|78557|4961|11138|389682504|4961 / 389687465|11138 / 89695|
|3|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|78545|4795|11150|389682670|4795 / 389687465|11150 / 89695|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 394111186 (0.000000%)|0 / 89895 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|89895|0|0|394111186|0 / 394111186|0 / 89895|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 389687465|0 / 89695|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999998|1849 / 394111186 (0.000469%)|0 / 89895 (0.000000%)|0.983812|0.989820|0.989869|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -178,7 +156,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.979846|1.000000|0.999995|0.999998|0.999995|0.000005|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -186,15 +164,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.983812|0.989820|0.995903|0.979846|0.989872|0.989869|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -202,7 +172,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|89895|1849|0|394109337|1849 / 394111186|0 / 89895|
|1|Radixor|ALL_CANDIDATES|89695|0|0|389687465|0 / 389687465|0 / 89695|
</details>
@@ -212,21 +182,21 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|707|1165|684|323|1.150326%|3|28405|
|Radixor|674|0|0|165|0.590953%|3|28087|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996305|1165 / 392820788 (0.000297%)|663 / 89740 (0.738801%)|0.988190|0.989843|0.989845|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.938074|6341 / 392820788 (0.001614%)|11113 / 89740 (12.383552%)|0.915093|0.900096|0.900410|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.938074|6341 / 392820788 (0.001614%)|11113 / 89740 (12.383552%)|0.915093|0.900096|0.900410|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.996482|0.000000%|0.703596%|
|2|SNOWBALL DANISH DIRECT|0.938010|0.001235%|12.396694%|
|3|SNOWBALL DANISH LUCENE FILTER|0.938010|0.001235%|12.396694%|
</div>
@@ -234,9 +204,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987090|0.992612|0.999997|0.996305|0.999995|0.000005|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.925372|0.876164|0.999984|0.938074|0.999956|0.000044|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.925372|0.876164|0.999984|0.938074|0.999956|0.000044|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.992964|1.000000|0.996482|0.999998|0.000002|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.942392|0.876033|0.999988|0.938010|0.999959|0.000041|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.942392|0.876033|0.999988|0.938010|0.999959|0.000041|
</details>
@@ -244,19 +214,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988190|0.989843|0.991503|0.979891|0.989847|0.989845|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.915093|0.900096|0.885583|0.818341|0.900432|0.900410|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.915093|0.900096|0.885583|0.818341|0.900432|0.900410|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989841|0.998463|0.998812|0.998637|0.998637|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.900074|0.994185|0.978644|0.986354|0.986354|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.900074|0.994185|0.978644|0.986354|0.986354|
|1|Radixor|PRIMARY_OUTPUT|0.998585|0.996470|0.994363|0.992964|0.996476|0.996475|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|0.928328|0.908002|0.888547|0.831505|0.908607|0.908587|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|0.928328|0.908002|0.888547|0.831505|0.908607|0.908587|
</details>
@@ -264,61 +224,39 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|89077|1165|663|392819623|1165 / 392820788|663 / 89740|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|78627|6341|11113|392814447|6341 / 392820788|11113 / 89740|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|78627|6341|11113|392814447|6341 / 392820788|11113 / 89740|
|1|Radixor|PRIMARY_OUTPUT|88910|0|630|388404335|0 / 388404335|630 / 89540|
|2|SNOWBALL DANISH DIRECT|PRIMARY_OUTPUT|78440|4795|11100|388399540|4795 / 388404335|11100 / 89540|
|3|SNOWBALL DANISH LUCENE FILTER|PRIMARY_OUTPUT|78440|4795|11100|388399540|4795 / 388404335|11100 / 89540|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 392820788 (0.000000%)|0 / 89740 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|89740|0|0|392820788|0 / 392820788|0 / 89740|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 388404335|0 / 89540|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999998|1849 / 392820788 (0.000471%)|0 / 89740 (0.000000%)|0.983784|0.989803|0.989852|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -326,7 +264,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.979812|1.000000|0.999995|0.999998|0.999995|0.000005|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -334,15 +272,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.983784|0.989803|0.995896|0.979812|0.989855|0.989852|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -350,7 +280,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|89740|1849|0|392818939|1849 / 392820788|0 / 89740|
|1|Radixor|ALL_CANDIDATES|89540|0|0|388404335|0 / 388404335|0 / 89540|
</details>
@@ -360,19 +290,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|663|1165|684|315|1.123676%|3|28351|
|Radixor|630|0|0|157|0.563229%|3|28033|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -381,16 +311,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `DA_DK`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

329
docs/benchmarks/languages/dutch.md
View File

@@ -8,13 +8,13 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `NL_NL` | 4,992 | 31,466 | 9,981 | 21,485 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `nl-nl-default` | `1.0.0` | `NL_NL` | 4,992 | 31,466 | 9,981 | 21,485 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **31,466**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **31,466**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
@@ -35,16 +35,26 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 15.954% | 8.992% | 30.939% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene SnowballFilter | 12.620% | 5.441% | 28.073% | Lucene TokenFilter integration path around the Snowball algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `radixor[DUTCH]` | 1.331 | 0.114 | 61.9 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 22.760 | 1.387 | 1059.3 | 17.105 | Benchmark-only Dutch Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Official Snowball direct | `snowballDirect[DUTCH]` | 4.146 | 0.291 | 193.0 | 3.116 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[DUTCH]` | 7.375 | 0.595 | 343.3 | 5.543 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `radixor[DUTCH]` | 1.410 | 0.139 | 65.6 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 24.183 | 2.889 | 1125.6 | 17.156 | Benchmark-only Dutch Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Official Snowball direct | `snowballDirect[DUTCH]` | 4.560 | 0.205 | 212.2 | 3.235 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[DUTCH]` | 7.762 | 0.262 | 361.3 | 5.506 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -58,30 +68,30 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `NL_NL` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `NL_NL` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/nl_nl/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `nl-nl-default`, loaded from classpath resource `org/egothor/stemmer/models/nl-nl-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.988661** among 4 deterministic stemmers. The runner-up is `SNOWBALL DUTCH DIRECT` at 0.727087, a difference of 0.261574. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.989040** among 4 deterministic stemmers. The runner-up is `SNOWBALL DUTCH DIRECT` at 0.730495, a difference of 0.258544. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.988733** among 4 deterministic stemmers. The runner-up is `SNOWBALL DUTCH DIRECT` at 0.727093, a difference of 0.261640. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.989114** among 4 deterministic stemmers. The runner-up is `SNOWBALL DUTCH DIRECT` at 0.730509, a difference of 0.258605. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988661|1214 / 350437960 (0.000346%)|1464 / 64566 (2.267447%)|0.980362|0.979221|0.979219|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.727087|4382 / 350437960 (0.001250%)|35241 / 64566 (54.581359%)|0.735353|0.596807|0.628557|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.643123|1333 / 350437960 (0.000380%)|46084 / 64566 (71.375027%)|0.642512|0.438061|0.516674|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.618497|1588 / 350437960 (0.000453%)|49264 / 64566 (76.300220%)|0.579068|0.375712|0.463333|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.988733|0.000000%|2.253364%|
|2|SNOWBALL DUTCH DIRECT|0.727093|0.000870%|54.580443%|
|3|HUNSPELL DUTCH LUCENE FILTER|0.642844|0.000104%|71.431010%|
|4|SNOWBALL DUTCH LUCENE FILTER|0.617975|0.000221%|76.404861%|
</div>
@@ -89,10 +99,10 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.981124|0.977326|0.999997|0.988661|0.999992|0.000008|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.869997|0.454186|0.999987|0.727087|0.999887|0.000113|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.932728|0.286250|0.999996|0.643123|0.999865|0.000135|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.905980|0.236998|0.999995|0.618497|0.999855|0.000145|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.977466|1.000000|0.988733|0.999996|0.000004|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.907391|0.454196|0.999991|0.727093|0.999889|0.000111|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.981029|0.285690|0.999999|0.642844|0.999865|0.000135|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.952453|0.235951|0.999998|0.617975|0.999854|0.000146|
</details>
@@ -100,21 +110,10 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.980362|0.979221|0.978083|0.959289|0.979223|0.979219|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.735353|0.596807|0.502190|0.425321|0.628602|0.628557|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.642512|0.438061|0.332316|0.280459|0.516714|0.516674|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.579068|0.375712|0.278062|0.231309|0.463374|0.463333|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.979217|0.997464|0.997003|0.997234|0.997234|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.596756|0.992815|0.917346|0.953590|0.953590|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.438012|0.996932|0.889026|0.939892|0.939892|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.375664|0.995828|0.888410|0.939057|0.939057|
|1|Radixor|PRIMARY_OUTPUT|0.995411|0.988605|0.981891|0.977466|0.988669|0.988667|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.756437|0.605372|0.504600|0.434074|0.641976|0.641934|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.659835|0.442513|0.332878|0.284120|0.529405|0.529368|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.592568|0.378209|0.277738|0.233204|0.474060|0.474022|
</details>
@@ -122,68 +121,43 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|63102|1214|1464|350436746|1214 / 350437960|1464 / 64566|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|29325|4382|35241|350433578|4382 / 350437960|35241 / 64566|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|18482|1333|46084|350436627|1333 / 350437960|46084 / 64566|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|15302|1588|49264|350436372|1588 / 350437960|49264 / 64566|
|1|Radixor|PRIMARY_OUTPUT|62985|0|1452|343168663|0 / 343168663|1452 / 64437|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|29267|2987|35170|343165676|2987 / 343168663|35170 / 64437|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|18409|356|46028|343168307|356 / 343168663|46028 / 64437|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|15204|759|49233|343167904|759 / 343168663|49233 / 64437|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 350437960 (0.000000%)|0 / 64566 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|0.665519|1164 / 350437960 (0.000332%)|43192 / 64566 (66.895889%)|0.690741|0.490770|0.560268|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL DUTCH LUCENE FILTER|0.000096%|66.975495%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|0.948354|0.331041|0.999997|0.665519|0.999873|0.000127|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|0.690741|0.490770|0.380588|0.325179|0.560307|0.560268|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|64566|0|0|350437960|0 / 350437960|0 / 64566|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|21374|1164|43192|350436796|1164 / 350437960|43192 / 64566|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL DUTCH LUCENE FILTER|330 / 343168663|43157 / 64437|
|Radixor|0 / 343168663|0 / 64437|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999996|2651 / 350437960 (0.000756%)|0 / 64566 (0.000000%)|0.968198|0.979884|0.980078|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.665518|1738 / 350437960 (0.000496%)|43192 / 64566 (66.895889%)|0.680640|0.487557|0.553265|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|HUNSPELL DUTCH LUCENE FILTER|0.665122|0.000147%|66.975495%|
</div>
@@ -191,8 +165,8 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.960561|1.000000|0.999992|0.999996|0.999992|0.000008|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.924801|0.331041|0.999995|0.665518|0.999872|0.000128|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.976909|0.330245|0.999999|0.665122|0.999873|0.000127|
</details>
@@ -200,17 +174,8 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.968198|0.979884|0.991855|0.960561|0.980082|0.980078|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.680640|0.487557|0.379812|0.322364|0.553306|0.553265|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.701991|0.493621|0.380638|0.327687|0.567996|0.567958|
</details>
@@ -218,8 +183,8 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|64566|2651|0|350435309|2651 / 350437960|0 / 64566|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|21374|1738|43192|350436222|1738 / 350437960|43192 / 64566|
|1|Radixor|ALL_CANDIDATES|64437|0|0|343168663|0 / 343168663|0 / 64437|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|21280|503|43157|343168160|503 / 343168663|43157 / 64437|
</details>
@@ -229,23 +194,23 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL DUTCH LUCENE FILTER|2892|169|405|1254|4.736186%|3|27763|
|Radixor|1464|1214|1437|572|2.160366%|3|27061|
|HUNSPELL DUTCH LUCENE FILTER|2871|26|147|1199|4.576161%|3|27429|
|Radixor|1452|0|0|296|1.129728%|3|26501|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989040|1214 / 329603856 (0.000368%)|1384 / 63147 (2.191711%)|0.980194|0.979401|0.979398|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.730495|4382 / 329603856 (0.001329%)|34036 / 63147 (53.899631%)|0.738412|0.602463|0.632953|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.645159|1310 / 329603856 (0.000397%)|44814 / 63147 (70.967742%)|0.646808|0.442880|0.520498|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.618546|1544 / 329603856 (0.000468%)|48175 / 63147 (76.290243%)|0.579362|0.375883|0.463566|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.989114|0.000000%|2.177156%|
|2|SNOWBALL DUTCH DIRECT|0.730509|0.000926%|53.897299%|
|3|HUNSPELL DUTCH LUCENE FILTER|0.644879|0.000103%|71.024152%|
|4|SNOWBALL DUTCH LUCENE FILTER|0.618013|0.000222%|76.397220%|
</div>
@@ -253,10 +218,10 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.980723|0.978083|0.999996|0.989040|0.999992|0.000008|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.869167|0.461004|0.999987|0.730495|0.999883|0.000117|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.933310|0.290323|0.999996|0.645159|0.999860|0.000140|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.906515|0.237098|0.999995|0.618546|0.999849|0.000151|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.978228|1.000000|0.989114|0.999996|0.000004|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.906773|0.461027|0.999991|0.730509|0.999885|0.000115|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.982090|0.289758|0.999999|0.644879|0.999860|0.000140|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.954134|0.236028|0.999998|0.618013|0.999849|0.000151|
</details>
@@ -264,21 +229,10 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.980194|0.979401|0.978610|0.959634|0.979402|0.979398|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.738412|0.602463|0.508789|0.431089|0.633000|0.632953|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.646808|0.442880|0.336718|0.284422|0.520539|0.520498|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.579362|0.375883|0.278182|0.231439|0.463608|0.463566|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.979397|0.997373|0.997139|0.997256|0.997256|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.602410|0.992557|0.918059|0.953856|0.953856|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.442829|0.996884|0.889061|0.939890|0.939890|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.375834|0.995817|0.887492|0.938539|0.938539|
|1|Radixor|PRIMARY_OUTPUT|0.995569|0.988994|0.982507|0.978228|0.989054|0.989052|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|0.759842|0.611269|0.511295|0.440164|0.646565|0.646521|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.664532|0.447489|0.337317|0.288235|0.533450|0.533411|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|0.593185|0.378440|0.277851|0.233380|0.474555|0.474515|
</details>
@@ -286,68 +240,43 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|61763|1214|1384|329602642|1214 / 329603856|1384 / 63147|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|29111|4382|34036|329599474|4382 / 329603856|34036 / 63147|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|18333|1310|44814|329602546|1310 / 329603856|44814 / 63147|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|14972|1544|48175|329602312|1544 / 329603856|48175 / 63147|
|1|Radixor|PRIMARY_OUTPUT|61646|0|1372|322555083|0 / 322555083|1372 / 63018|
|2|SNOWBALL DUTCH DIRECT|PRIMARY_OUTPUT|29053|2987|33965|322552096|2987 / 322555083|33965 / 63018|
|3|HUNSPELL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|18260|333|44758|322554750|333 / 322555083|44758 / 63018|
|4|SNOWBALL DUTCH LUCENE FILTER|PRIMARY_OUTPUT|14874|715|48144|322554368|715 / 322555083|48144 / 63018|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 329603856 (0.000000%)|0 / 63147 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|0.667956|1141 / 329603856 (0.000346%)|41935 / 63147 (66.408539%)|0.695206|0.496187|0.564555|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL DUTCH LUCENE FILTER|0.000095%|66.488940%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|0.948955|0.335915|0.999997|0.667956|0.999869|0.000131|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|0.695206|0.496187|0.385755|0.329953|0.564595|0.564555|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|63147|0|0|329603856|0 / 329603856|0 / 63147|
|2|HUNSPELL DUTCH LUCENE FILTER|ANY_CANDIDATE|21212|1141|41935|329602715|1141 / 329603856|41935 / 63147|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL DUTCH LUCENE FILTER|307 / 322555083|41900 / 63018|
|Radixor|0 / 322555083|0 / 63018|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999996|2651 / 329603856 (0.000804%)|0 / 63147 (0.000000%)|0.967506|0.979441|0.979644|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.667955|1712 / 329603856 (0.000519%)|41935 / 63147 (66.408539%)|0.684952|0.492895|0.557477|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|HUNSPELL DUTCH LUCENE FILTER|0.667555|0.000148%|66.488940%|
</div>
@@ -355,8 +284,8 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.959710|1.000000|0.999992|0.999996|0.999992|0.000008|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.925318|0.335915|0.999995|0.667955|0.999868|0.000132|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.977912|0.335111|0.999999|0.667555|0.999869|0.000131|
</details>
@@ -364,17 +293,8 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.967506|0.979441|0.991674|0.959710|0.979648|0.979644|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.684952|0.492895|0.384956|0.327048|0.557519|0.557477|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|0.706770|0.499167|0.385834|0.332593|0.572458|0.572419|
</details>
@@ -382,8 +302,8 @@ This mode contains **8 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|63147|2651|0|329601205|2651 / 329603856|0 / 63147|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|21212|1712|41935|329602144|1712 / 329603856|41935 / 63147|
|1|Radixor|ALL_CANDIDATES|63018|0|0|322555083|0 / 322555083|0 / 63018|
|2|HUNSPELL DUTCH LUCENE FILTER|ALL_CANDIDATES|21118|477|41900|322554606|477 / 322555083|41900 / 63018|
</details>
@@ -393,20 +313,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL DUTCH LUCENE FILTER|2879|169|402|1186|4.618740%|3|26896|
|Radixor|1384|1214|1437|549|2.138017%|3|26239|
|HUNSPELL DUTCH LUCENE FILTER|2858|26|144|1131|4.452405%|3|26562|
|Radixor|1372|0|0|273|1.074719%|3|25679|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -415,16 +335,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `NL_NL`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

477
docs/benchmarks/languages/english.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `US_UK` | 396,939 | 1,004,374 | 793,874 | 210,500 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `us-uk-default` | `1.0.0` | `US_UK` | 396,939 | 1,004,374 | 793,874 | 210,500 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **1,004,374**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **1,004,374**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 28 | 0.003% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 22,493 | 2.240% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 186,764 | 18.595% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 795,024 | 79.156% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 65 | 0.006% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 73 | 0.007% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 22,481 | 2.238% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 202,637 | 20.175% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 779,106 | 77.571% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 77 | 0.008% |
## Accuracy
@@ -42,23 +42,31 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Snowball original Porter | 39.529% | 46.179% | 37.766% | Classic Porter rule-based suffix stemmer. |
| Paice/Husk Lancaster | 28.055% | 37.039% | 25.673% | Aggressive Paice/Husk rule stemmer that often produces shorter stems. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `radixorUsUkProfiPreferredStem` | 21.987 | 8.707 | 104.5 | 1.000 | Full dictionary patch-command stemmer using compiled patch commands. |
| Lucene EnglishPossessiveFilter | `luceneEnglishPossessiveFilter` | 24.539 | 1.515 | 116.6 | 1.116 | Possessive-ending remover only; not a full stemmer. |
| Lucene EnglishMinimalStemFilter | `luceneEnglishMinimalStemFilter` | 22.702 | 1.195 | 107.8 | 1.032 | Narrow plural reduction filter; not a full stemmer. |
| Lucene PorterStemmer direct copy | `lucenePorterStemmerCopied` | 24.696 | 13.235 | 117.3 | 1.123 | Benchmark-only generated copy of Lucene package-private Porter implementation. |
| OpenNLP PorterStemmer | `opennlpPorterStemmer` | 23.121 | 12.528 | 109.8 | 1.052 | Apache OpenNLP Porter implementation. |
| Snowball original Porter | `snowballOriginalPorter` | 38.904 | 10.353 | 184.8 | 1.769 | Classic Porter suffix-rule stemmer; historical English baseline, not a dictionary-equivalent stemmer. |
| Lucene PorterStemFilter | `lucenePorterStemFilter` | 37.021 | 1.196 | 175.9 | 1.684 | Lucene TokenFilter integration path for Porter; includes TokenStream overhead. |
| Lucene KStemFilter | `luceneKStemFilter` | 51.640 | 2.591 | 245.3 | 2.349 | Krovetz-style English TokenFilter; broader than minimal suffix filters. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 79.785 | 1.347 | 379.0 | 3.629 | Benchmark-only English Hunspell comparison using the benchmark Hunspell corpus. |
| Snowball English / Porter2 | `snowballEnglishPorter2` | 52.437 | 0.773 | 249.1 | 2.385 | Porter2 suffix-rule stemmer, distinct from original Porter. |
| Paice/Husk Lancaster | `paiceHuskLancaster` | 141.556 | 12.324 | 672.5 | 6.438 | Aggressive rule-based English stemmer. |
| Radixor | `radixorUsUkProfiPreferredStem` | 17.489 | 1.380 | 83.1 | 1.000 | Full dictionary patch-command stemmer using compiled patch commands. |
| Lucene EnglishPossessiveFilter | `luceneEnglishPossessiveFilter` | 17.151 | 0.215 | 81.5 | 0.981 | Possessive-ending remover only; not a full stemmer. |
| Lucene EnglishMinimalStemFilter | `luceneEnglishMinimalStemFilter` | 18.522 | 0.152 | 88.0 | 1.059 | Narrow plural reduction filter; not a full stemmer. |
| Lucene PorterStemmer direct copy | `lucenePorterStemmerCopied` | 17.651 | 0.129 | 83.9 | 1.009 | Benchmark-only generated copy of Lucene package-private Porter implementation. |
| OpenNLP PorterStemmer | `opennlpPorterStemmer` | 17.681 | 0.139 | 84.0 | 1.011 | Apache OpenNLP Porter implementation. |
| Snowball original Porter | `snowballOriginalPorter` | 33.290 | 1.916 | 158.1 | 1.904 | Classic Porter suffix-rule stemmer; historical English baseline, not a dictionary-equivalent stemmer. |
| Lucene PorterStemFilter | `lucenePorterStemFilter` | 32.408 | 0.412 | 154.0 | 1.853 | Lucene TokenFilter integration path for Porter; includes TokenStream overhead. |
| Lucene KStemFilter | `luceneKStemFilter` | 45.877 | 0.425 | 217.9 | 2.623 | Krovetz-style English TokenFilter; broader than minimal suffix filters. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 76.852 | 1.028 | 365.1 | 4.394 | Benchmark-only English Hunspell comparison using the benchmark Hunspell corpus. |
| Snowball English / Porter2 | `snowballEnglishPorter2` | 46.568 | 2.414 | 221.2 | 2.663 | Porter2 suffix-rule stemmer, distinct from original Porter. |
| Paice/Husk Lancaster | `paiceHuskLancaster` | 144.951 | 2.710 | 688.6 | 8.288 | Aggressive rule-based English stemmer. |
## Interpretation Notes
@@ -72,37 +80,37 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `US_UK` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `US_UK` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/us_uk/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `us-uk-default`, loaded from classpath resource `org/egothor/stemmer/models/us-uk-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.965159** among 11 deterministic stemmers. The runner-up is `ENGLISH LUCENE PORTER COPIED` at 0.954627, a difference of 0.010533. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.965820** among 11 deterministic stemmers. The runner-up is `ENGLISH LUCENE PORTER COPIED` at 0.954900, a difference of 0.010920. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.965537** among 11 deterministic stemmers. The runner-up is `ENGLISH LUCENE PORTER COPIED` at 0.954796, a difference of 0.010741. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.966202** among 11 deterministic stemmers. The runner-up is `ENGLISH LUCENE PORTER COPIED` at 0.955064, a difference of 0.011139. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.965159|1149886 / 184490451771 (0.000623%)|21869 / 313870 (6.967534%)|0.240076|0.332621|0.434052|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.954627|1557406 / 184490451771 (0.000844%)|28480 / 313870 (9.073820%)|0.185679|0.264659|0.375252|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.954627|1557406 / 184490451771 (0.000844%)|28480 / 313870 (9.073820%)|0.185679|0.264659|0.375252|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.954627|1557406 / 184490451771 (0.000844%)|28480 / 313870 (9.073820%)|0.185679|0.264659|0.375252|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.954537|1566711 / 184490451771 (0.000849%)|28536 / 313870 (9.091662%)|0.184753|0.263477|0.374240|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.954490|1555293 / 184490451771 (0.000843%)|28566 / 313870 (9.101220%)|0.185835|0.264849|0.375363|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.952394|3062661 / 184490451771 (0.001660%)|29879 / 313870 (9.519546%)|0.103643|0.155164|0.277089|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.878441|1368501 / 184490451771 (0.000742%)|76305 / 313870 (24.311020%)|0.176284|0.247472|0.334598|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.718599|1122264 / 184490451771 (0.000608%)|176645 / 313870 (56.279670%)|0.128204|0.174436|0.218251|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.573277|1981986 / 184490451771 (0.001074%)|267868 / 313870 (85.343614%)|0.027298|0.039287|0.057655|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.500008|1115154 / 184490451771 (0.000604%)|313863 / 313870 (99.997770%)|0.000007|0.000010|0.000009|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.965537|&lt;0.000001%|6.892502%|
|2|ENGLISH LUCENE PORTER COPIED|0.954796|0.000207%|9.040545%|
|3|ENGLISH LUCENE PORTER FILTER|0.954796|0.000207%|9.040545%|
|4|ENGLISH OPENNLP PORTER|0.954796|0.000207%|9.040545%|
|5|ENGLISH SNOWBALL PORTER2|0.954708|0.000212%|9.058097%|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|0.954659|0.000206%|9.067990%|
|7|ENGLISH PAICE HUSK LANCASTER|0.952535|0.000960%|9.492110%|
|8|ENGLISH LUCENE KSTEM FILTER|0.878645|0.000110%|24.270875%|
|9|ENGLISH LUCENE MINIMAL FILTER|0.718958|0.000001%|56.208454%|
|10|HUNSPELL ENGLISH LUCENE FILTER|0.573139|0.000012%|85.372182%|
|11|ENGLISH LUCENE POSSESSIVE FILTER|0.500011|&lt;0.000001%|99.997766%|
</div>
@@ -110,17 +118,17 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.202513|0.930325|0.999994|0.965159|0.999994|0.000006|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.154868|0.909262|0.999992|0.954627|0.999991|0.000009|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.154868|0.909262|0.999992|0.954627|0.999991|0.000009|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.154868|0.909262|0.999992|0.954627|0.999991|0.000009|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.154064|0.909083|0.999992|0.954537|0.999991|0.000009|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.155006|0.908988|0.999992|0.954490|0.999991|0.000009|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.084858|0.904805|0.999983|0.952394|0.999983|0.000017|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.147917|0.756890|0.999993|0.878441|0.999992|0.000008|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.108953|0.437203|0.999994|0.718599|0.999993|0.000007|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.022684|0.146564|0.999989|0.573277|0.999988|0.000012|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000006|0.000022|0.999994|0.500008|0.999992|0.000008|
|1|Radixor|PRIMARY_OUTPUT|0.999990|0.931075|1.000000|0.965537|1.000000|0.000000|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.440121|0.909595|0.999998|0.954796|0.999998|0.000002|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.440121|0.909595|0.999998|0.954796|0.999998|0.000002|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.440121|0.909595|0.999998|0.954796|0.999998|0.000002|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.434174|0.909419|0.999998|0.954708|0.999998|0.000002|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.441440|0.909320|0.999998|0.954659|0.999998|0.000002|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.144284|0.905079|0.999990|0.952535|0.999990|0.000010|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.551014|0.757291|0.999999|0.878645|0.999998|0.000002|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.989894|0.437915|1.000000|0.718958|0.999999|0.000001|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.681277|0.146278|1.000000|0.573139|0.999998|0.000002|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.148936|0.000022|1.000000|0.500011|0.999998|0.000002|
</details>
@@ -128,35 +136,17 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.240076|0.332621|0.541270|0.199487|0.434054|0.434052|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.185679|0.264659|0.460563|0.152511|0.375254|0.375252|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.185679|0.264659|0.460563|0.152511|0.375254|0.375252|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.185679|0.264659|0.460563|0.152511|0.375254|0.375252|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.184753|0.263477|0.459102|0.151727|0.374242|0.374240|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.185835|0.264849|0.460751|0.152637|0.375365|0.375363|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.103643|0.155164|0.308543|0.084107|0.277092|0.277089|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.176284|0.247472|0.415099|0.141208|0.334600|0.334598|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.128204|0.174436|0.272816|0.095552|0.218253|0.218251|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.027298|0.039287|0.070051|0.020037|0.057659|0.057655|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000007|0.000010|0.000015|0.000005|0.000012|0.000009|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.332619|0.994215|0.997770|0.995989|0.995989|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.264656|0.969648|0.997199|0.983231|0.983231|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.264656|0.969648|0.997199|0.983231|0.983231|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.264656|0.969648|0.997199|0.983231|0.983231|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.263474|0.969037|0.997182|0.982908|0.982908|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.264847|0.969891|0.997193|0.983353|0.983353|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.155162|0.937768|0.996600|0.966289|0.966289|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.247470|0.980687|0.992108|0.986364|0.986364|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.174433|0.995202|0.981174|0.988138|0.988138|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.039284|0.993096|0.963677|0.978166|0.978166|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000007|0.995789|0.958019|0.976539|0.976539|
|1|Radixor|PRIMARY_OUTPUT|0.985403|0.964303|0.944087|0.931066|0.964917|0.964917|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.490783|0.593208|0.749662|0.421675|0.632717|0.632716|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.490783|0.593208|0.749662|0.421675|0.632717|0.632716|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.490783|0.593208|0.749662|0.421675|0.632717|0.632716|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.484849|0.587747|0.746086|0.416176|0.628368|0.628367|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.492079|0.594348|0.750277|0.422827|0.633570|0.633569|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.173443|0.248891|0.440518|0.142133|0.361370|0.361368|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.582762|0.637891|0.704541|0.468312|0.645971|0.645970|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.790591|0.607210|0.492883|0.435966|0.658399|0.658399|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.393465|0.240844|0.173533|0.136909|0.315683|0.315683|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000112|0.000045|0.000028|0.000022|0.001824|0.001824|
</details>
@@ -164,75 +154,50 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|292001|1149886|21869|184489301885|1149886 / 184490451771|21869 / 313870|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|285390|1557406|28480|184488894365|1557406 / 184490451771|28480 / 313870|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|285390|1557406|28480|184488894365|1557406 / 184490451771|28480 / 313870|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|285390|1557406|28480|184488894365|1557406 / 184490451771|28480 / 313870|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|285334|1566711|28536|184488885060|1566711 / 184490451771|28536 / 313870|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|285304|1555293|28566|184488896478|1555293 / 184490451771|28566 / 313870|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|283991|3062661|29879|184487389110|3062661 / 184490451771|29879 / 313870|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|237565|1368501|76305|184489083270|1368501 / 184490451771|76305 / 313870|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|137225|1122264|176645|184489329507|1122264 / 184490451771|176645 / 313870|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|46002|1981986|267868|184488469785|1981986 / 184490451771|267868 / 313870|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|7|1115154|313863|184489336617|1115154 / 184490451771|313863 / 313870|
|1|Radixor|PRIMARY_OUTPUT|291757|3|21598|175199424127|3 / 175199424130|21598 / 313355|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|285026|362583|28329|175199061547|362583 / 175199424130|28329 / 313355|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|285026|362583|28329|175199061547|362583 / 175199424130|28329 / 313355|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|285026|362583|28329|175199061547|362583 / 175199424130|28329 / 313355|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|284971|371381|28384|175199052749|371381 / 175199424130|28384 / 313355|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|284940|360538|28415|175199063592|360538 / 175199424130|28415 / 313355|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|283611|1682034|29744|175197742096|1682034 / 175199424130|29744 / 313355|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|237301|193361|76054|175199230769|193361 / 175199424130|76054 / 313355|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|137223|1401|176132|175199422729|1401 / 175199424130|176132 / 313355|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|45837|21444|267518|175199402686|21444 / 175199424130|267518 / 313355|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|7|40|313348|175199424090|40 / 175199424130|313348 / 313355|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999976|12 / 184490451771 (0.000000%)|15 / 313870 (0.004779%)|0.999960|0.999957|0.999957|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|0.581603|1978852 / 184490451771 (0.001073%)|262641 / 313870 (83.678274%)|0.030370|0.043712|0.064174|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.004787%|
|HUNSPELL ENGLISH LUCENE FILTER|0.000012%|83.719424%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999962|0.999952|1.000000|0.999976|1.000000|0.000000|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|0.025235|0.163217|0.999989|0.581603|0.999988|0.000012|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999960|0.999957|0.999954|0.999914|0.999957|0.999957|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|0.030370|0.043712|0.077961|0.022344|0.064178|0.064174|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|313855|12|15|184490451759|12 / 184490451771|15 / 313870|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|51229|1978852|262641|184488472919|1978852 / 184490451771|262641 / 313870|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 175199424130|15 / 313355|
|HUNSPELL ENGLISH LUCENE FILTER|20367 / 175199424130|262339 / 313355|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999945|11482166 / 184490451771 (0.006224%)|15 / 313870 (0.004779%)|0.033039|0.051834|0.163107|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.581603|2008917 / 184490451771 (0.001089%)|262641 / 313870 (83.678274%)|0.029943|0.043158|0.063704|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.999976|&lt;0.000001%|0.004787%|
|2|HUNSPELL ENGLISH LUCENE FILTER|0.581403|0.000022%|83.719424%|
</div>
@@ -240,8 +205,8 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.026607|0.999952|0.999938|0.999945|0.999938|0.000062|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.024867|0.163217|0.999989|0.581603|0.999988|0.000012|
|1|Radixor|ALL_CANDIDATES|0.999825|0.999952|1.000000|0.999976|1.000000|0.000000|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.568132|0.162806|1.000000|0.581403|0.999998|0.000002|
</details>
@@ -249,17 +214,8 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.033039|0.051834|0.120237|0.026607|0.163112|0.163107|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.029943|0.043158|0.077254|0.022055|0.063708|0.063704|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999850|0.999888|0.999927|0.999777|0.999888|0.999888|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.379279|0.253086|0.189902|0.144876|0.304130|0.304130|
</details>
@@ -267,8 +223,8 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|313855|11482166|15|184478969605|11482166 / 184490451771|15 / 313870|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|51229|2008917|262641|184488442854|2008917 / 184490451771|262641 / 313870|
|1|Radixor|ALL_CANDIDATES|313340|55|15|175199424075|55 / 175199424130|15 / 313355|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|51016|38780|262339|175199385350|38780 / 175199424130|262339 / 313355|
</details>
@@ -278,30 +234,30 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|21854|1149874|10332280|29208|4.808384%|1355|2838145|
|HUNSPELL ENGLISH LUCENE FILTER|5227|3134|26931|6837|1.125545%|4|614296|
|Radixor|21583|3|52|13718|2.317441%|1355|607918|
|HUNSPELL ENGLISH LUCENE FILTER|5179|1077|17336|5736|0.969007%|4|597698|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.965820|1148489 / 170474840204 (0.000674%)|21319 / 311891 (6.835401%)|0.239424|0.331902|0.433722|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.954900|1552702 / 170474840204 (0.000911%)|28130 / 311891 (9.019177%)|0.185277|0.264166|0.374937|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.954900|1552702 / 170474840204 (0.000911%)|28130 / 311891 (9.019177%)|0.185277|0.264166|0.374937|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.954900|1552702 / 170474840204 (0.000911%)|28130 / 311891 (9.019177%)|0.185277|0.264166|0.374937|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.954850|1561891 / 170474840204 (0.000916%)|28161 / 311891 (9.029116%)|0.184375|0.263016|0.373964|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.954762|1550615 / 170474840204 (0.000910%)|28216 / 311891 (9.046750%)|0.185431|0.264353|0.375045|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.952710|3045870 / 170474840204 (0.001787%)|29493 / 311891 (9.456188%)|0.103633|0.155157|0.277170|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.880820|1367069 / 170474840204 (0.000802%)|74340 / 311891 (23.835250%)|0.176477|0.247899|0.335789|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.719516|1120871 / 170474840204 (0.000657%)|174959 / 311891 (56.096200%)|0.128139|0.174470|0.218621|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.573619|1978041 / 170474840204 (0.001160%)|265965 / 311891 (85.274984%)|0.027312|0.039323|0.057799|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.500005|1113773 / 170474840204 (0.000653%)|311886 / 311891 (99.998397%)|0.000005|0.000007|0.000005|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.966202|&lt;0.000001%|6.759543%|
|2|ENGLISH LUCENE PORTER COPIED|0.955064|0.000222%|8.987032%|
|3|ENGLISH LUCENE PORTER FILTER|0.955064|0.000222%|8.987032%|
|4|ENGLISH OPENNLP PORTER|0.955064|0.000222%|8.987032%|
|5|ENGLISH SNOWBALL PORTER2|0.955016|0.000228%|8.996666%|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|0.954926|0.000221%|9.014651%|
|7|ENGLISH PAICE HUSK LANCASTER|0.952850|0.001032%|9.428933%|
|8|ENGLISH LUCENE KSTEM FILTER|0.881028|0.000120%|23.794246%|
|9|ENGLISH LUCENE MINIMAL FILTER|0.719875|0.000001%|56.025075%|
|10|HUNSPELL ENGLISH LUCENE FILTER|0.573484|0.000012%|85.303261%|
|11|ENGLISH LUCENE POSSESSIVE FILTER|0.500008|&lt;0.000001%|99.998394%|
</div>
@@ -309,17 +265,17 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.201918|0.931646|0.999993|0.965820|0.999993|0.000007|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.154515|0.909808|0.999991|0.954900|0.999991|0.000009|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.154515|0.909808|0.999991|0.954900|0.999991|0.000009|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.154515|0.909808|0.999991|0.954900|0.999991|0.000009|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.153731|0.909709|0.999991|0.954850|0.999991|0.000009|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.154651|0.909532|0.999991|0.954762|0.999991|0.000009|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.084848|0.905438|0.999982|0.952710|0.999982|0.000018|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.148042|0.761647|0.999992|0.880820|0.999992|0.000008|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.108866|0.439038|0.999993|0.719516|0.999992|0.000008|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.022691|0.147250|0.999988|0.573619|0.999987|0.000013|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000004|0.000016|0.999993|0.500005|0.999992|0.000008|
|1|Radixor|PRIMARY_OUTPUT|0.999990|0.932405|1.000000|0.966202|1.000000|0.000000|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.440920|0.910130|0.999998|0.955064|0.999998|0.000002|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.440920|0.910130|0.999998|0.955064|0.999998|0.000002|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.440920|0.910130|0.999998|0.955064|0.999998|0.000002|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.435017|0.910033|0.999998|0.955016|0.999998|0.000002|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.442235|0.909853|0.999998|0.954926|0.999998|0.000002|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.144700|0.905711|0.999990|0.952850|0.999990|0.000010|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.551013|0.762058|0.999999|0.881028|0.999998|0.000002|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.989965|0.439749|1.000000|0.719875|0.999999|0.000001|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.700136|0.146967|1.000000|0.573484|0.999998|0.000002|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.121951|0.000016|1.000000|0.500008|0.999998|0.000002|
</details>
@@ -327,35 +283,17 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.239424|0.331902|0.540775|0.198970|0.433723|0.433722|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.185277|0.264166|0.460049|0.152184|0.374939|0.374937|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.185277|0.264166|0.460049|0.152184|0.374939|0.374937|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.185277|0.264166|0.460049|0.152184|0.374939|0.374937|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.184375|0.263016|0.458637|0.151421|0.373966|0.373964|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.185431|0.264353|0.460234|0.152308|0.375047|0.375045|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.103633|0.155157|0.308576|0.084103|0.277173|0.277170|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.176477|0.247899|0.416437|0.141487|0.335791|0.335789|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.128139|0.174470|0.273277|0.095572|0.218624|0.218621|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.027312|0.039323|0.070190|0.020056|0.057804|0.057799|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000005|0.000007|0.000011|0.000004|0.000008|0.000005|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.331900|0.993959|0.997731|0.995842|0.995842|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.264164|0.968645|0.997109|0.982671|0.982671|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.264164|0.968645|0.997109|0.982671|0.982671|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.264164|0.968645|0.997109|0.982671|0.982671|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.263014|0.968020|0.997096|0.982343|0.982343|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.264351|0.968894|0.997102|0.982795|0.982795|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.155154|0.936077|0.996487|0.965338|0.965338|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.247897|0.979822|0.991991|0.985869|0.985869|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.174467|0.994994|0.980520|0.987704|0.987704|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.039320|0.993066|0.962317|0.977450|0.977450|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000004|0.995605|0.956423|0.975621|0.975621|
|1|Radixor|PRIMARY_OUTPUT|0.985700|0.965015|0.945181|0.932396|0.965606|0.965606|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|0.491609|0.594049|0.750417|0.422524|0.633478|0.633477|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|0.491609|0.594049|0.750417|0.422524|0.633478|0.633477|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|0.491609|0.594049|0.750417|0.422524|0.633478|0.633477|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|0.485725|0.588647|0.746915|0.417080|0.629190|0.629189|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|0.492900|0.595181|0.751027|0.423671|0.634326|0.634325|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|0.173928|0.249533|0.441413|0.142553|0.362017|0.362015|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|0.583322|0.639575|0.707836|0.470129|0.648000|0.647999|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|0.791820|0.608984|0.494744|0.437798|0.659800|0.659800|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|0.399444|0.242939|0.174549|0.138264|0.320776|0.320775|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|0.000080|0.000032|0.000020|0.000016|0.001399|0.001399|
</details>
@@ -363,75 +301,50 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|290572|1148489|21319|170473691715|1148489 / 170474840204|21319 / 311891|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|283761|1552702|28130|170473287502|1552702 / 170474840204|28130 / 311891|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|283761|1552702|28130|170473287502|1552702 / 170474840204|28130 / 311891|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|283761|1552702|28130|170473287502|1552702 / 170474840204|28130 / 311891|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|283730|1561891|28161|170473278313|1561891 / 170474840204|28161 / 311891|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|283675|1550615|28216|170473289589|1550615 / 170474840204|28216 / 311891|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|282398|3045870|29493|170471794334|3045870 / 170474840204|29493 / 311891|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|237551|1367069|74340|170473473135|1367069 / 170474840204|74340 / 311891|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|136932|1120871|174959|170473719333|1120871 / 170474840204|174959 / 311891|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|45926|1978041|265965|170472862163|1978041 / 170474840204|265965 / 311891|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|5|1113773|311886|170473726431|1113773 / 170474840204|311886 / 311891|
|1|Radixor|PRIMARY_OUTPUT|290334|3|21048|161561989635|3 / 161561989638|21048 / 311382|
|2|ENGLISH LUCENE PORTER COPIED|PRIMARY_OUTPUT|283398|359344|27984|161561630294|359344 / 161561989638|27984 / 311382|
|3|ENGLISH LUCENE PORTER FILTER|PRIMARY_OUTPUT|283398|359344|27984|161561630294|359344 / 161561989638|27984 / 311382|
|4|ENGLISH OPENNLP PORTER|PRIMARY_OUTPUT|283398|359344|27984|161561630294|359344 / 161561989638|27984 / 311382|
|5|ENGLISH SNOWBALL PORTER2|PRIMARY_OUTPUT|283368|368027|28014|161561621611|368027 / 161561989638|28014 / 311382|
|6|ENGLISH SNOWBALL ORIGINAL PORTER|PRIMARY_OUTPUT|283312|357325|28070|161561632313|357325 / 161561989638|28070 / 311382|
|7|ENGLISH PAICE HUSK LANCASTER|PRIMARY_OUTPUT|282022|1666990|29360|161560322648|1666990 / 161561989638|29360 / 311382|
|8|ENGLISH LUCENE KSTEM FILTER|PRIMARY_OUTPUT|237291|193354|74091|161561796284|193354 / 161561989638|74091 / 311382|
|9|ENGLISH LUCENE MINIMAL FILTER|PRIMARY_OUTPUT|136930|1388|174452|161561988250|1388 / 161561989638|174452 / 311382|
|10|HUNSPELL ENGLISH LUCENE FILTER|PRIMARY_OUTPUT|45763|19600|265619|161561970038|19600 / 161561989638|265619 / 311382|
|11|ENGLISH LUCENE POSSESSIVE FILTER|PRIMARY_OUTPUT|5|36|311377|161561989602|36 / 161561989638|311377 / 311382|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 170474840204 (0.000000%)|0 / 311891 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|0.581994|1974950 / 170474840204 (0.001158%)|260741 / 311891 (83.600040%)|0.030387|0.043756|0.064341|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
|HUNSPELL ENGLISH LUCENE FILTER|0.000011%|83.640994%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|0.025246|0.164000|0.999988|0.581994|0.999987|0.000013|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|0.030387|0.043756|0.078123|0.022367|0.064345|0.064341|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|311891|0|0|170474840204|0 / 170474840204|0 / 311891|
|2|HUNSPELL ENGLISH LUCENE FILTER|ANY_CANDIDATE|51150|1974950|260741|170472865254|1974950 / 170474840204|260741 / 311891|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 161561989638|0 / 311382|
|HUNSPELL ENGLISH LUCENE FILTER|18564 / 161561989638|260443 / 311382|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999966|11470018 / 170474840204 (0.006728%)|0 / 311891 (0.000000%)|0.032872|0.051579|0.162697|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.581994|2004598 / 170474840204 (0.001176%)|260741 / 311891 (83.600040%)|0.029965|0.043208|0.063875|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
|2|HUNSPELL ENGLISH LUCENE FILTER|0.581795|0.000023%|83.640994%|
</div>
@@ -439,8 +352,8 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.026472|1.000000|0.999933|0.999966|0.999933|0.000067|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.024881|0.164000|0.999988|0.581994|0.999987|0.000013|
|1|Radixor|ALL_CANDIDATES|0.999952|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.581828|0.163590|1.000000|0.581795|0.999998|0.000002|
</details>
@@ -448,17 +361,8 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.032872|0.051579|0.119687|0.026472|0.162702|0.162697|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.029965|0.043208|0.077422|0.022081|0.063879|0.063875|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999961|0.999976|0.999990|0.999952|0.999976|0.999976|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|0.384979|0.255377|0.191058|0.146379|0.308515|0.308514|
</details>
@@ -466,8 +370,8 @@ This mode contains **15 result rows**, **11 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|311891|11470018|0|170463370186|11470018 / 170474840204|0 / 311891|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|51150|2004598|260741|170472835606|2004598 / 170474840204|260741 / 311891|
|1|Radixor|ALL_CANDIDATES|311382|15|0|161561989623|15 / 161561989638|0 / 311382|
|2|HUNSPELL ENGLISH LUCENE FILTER|ALL_CANDIDATES|50939|36611|260443|161561953027|36611 / 161561989638|260443 / 311382|
</details>
@@ -477,20 +381,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|21319|1148489|10321529|28826|4.936720%|1355|2812871|
|HUNSPELL ENGLISH LUCENE FILTER|5224|3091|26557|6786|1.162165%|4|590716|
|Radixor|21048|3|12|13357|2.349760%|1355|584042|
|HUNSPELL ENGLISH LUCENE FILTER|5176|1036|17011|5685|1.000104%|4|574142|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -499,16 +403,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `US_UK`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

301
docs/benchmarks/languages/finnish.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `FI_FI` | 57,027 | 1,865,215 | 110,525 | 1,754,690 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `fi-fi-default` | `1.0.0` | `FI_FI` | 57,027 | 1,865,215 | 110,525 | 1,754,690 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **1,865,215**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **1,865,215**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 745 | 0.040% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 1,176,003 | 63.049% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 565,585 | 30.323% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 116,946 | 6.270% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 5,936 | 0.318% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 1,117 | 0.060% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 1,175,880 | 63.043% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 570,130 | 30.566% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 112,029 | 6.006% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 6,059 | 0.325% |
## Accuracy
@@ -35,16 +35,24 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 10.991% | 10.268% | 22.471% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene FinnishLightStemFilter | 4.351% | 4.294% | 5.264% | Light suffix stemmer; intentionally narrower than a dictionary-derived stemmer. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `finnishRadixor` | 308.076 | 15.529 | 175.6 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene FinnishLightStemFilter | `finnishLuceneFinnishLightStemFilter` | 175.250 | 46.995 | 99.9 | 0.569 | Light Finnish suffix stemmer. |
| Official Snowball direct | `snowballDirect[FINNISH]` | 264.652 | 63.054 | 150.8 | 0.859 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[FINNISH]` | 374.883 | 238.157 | 213.6 | 1.217 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `finnishRadixor` | 289.539 | 4.136 | 165.0 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene FinnishLightStemFilter | `finnishLuceneFinnishLightStemFilter` | 175.789 | 4.827 | 100.2 | 0.607 | Light Finnish suffix stemmer. |
| Official Snowball direct | `snowballDirect[FINNISH]` | 259.889 | 8.924 | 148.1 | 0.898 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[FINNISH]` | 332.524 | 9.490 | 189.5 | 1.148 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -58,30 +66,30 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `FI_FI` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `FI_FI` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/fi_fi/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `fi-fi-default`, loaded from classpath resource `org/egothor/stemmer/models/fi-fi-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.984594** among 4 deterministic stemmers. The runner-up is `SNOWBALL FINNISH LUCENE FILTER` at 0.740353, a difference of 0.244242. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.988068** among 4 deterministic stemmers. The runner-up is `SNOWBALL FINNISH DIRECT` at 0.738400, a difference of 0.249668. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.984838** among 4 deterministic stemmers. The runner-up is `SNOWBALL FINNISH LUCENE FILTER` at 0.740279, a difference of 0.244559. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.988242** among 4 deterministic stemmers. The runner-up is `SNOWBALL FINNISH DIRECT` at 0.738344, a difference of 0.249898. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.984594|731279 / 1641126814491 (0.000045%)|971268 / 31523695 (3.081073%)|0.975128|0.972893|0.972899|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.740353|1922153 / 1641126814491 (0.000117%)|16370057 / 31523695 (51.929372%)|0.758996|0.623613|0.653138|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.739729|1544812 / 1641126814491 (0.000094%)|16409363 / 31523695 (52.054060%)|0.769880|0.627374|0.659540|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.695969|2223150 / 1641126814491 (0.000135%)|19168306 / 31523695 (60.806025%)|0.687649|0.536000|0.576338|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.984838|&lt;0.000001%|3.032474%|
|2|SNOWBALL FINNISH LUCENE FILTER|0.740279|0.000081%|51.944179%|
|3|SNOWBALL FINNISH DIRECT|0.739671|0.000060%|52.065724%|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|0.695725|0.000094%|60.854936%|
</div>
@@ -89,10 +97,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.976624|0.969189|1.000000|0.984594|0.999999|0.000001|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.887434|0.480706|0.999999|0.740353|0.999989|0.000011|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.907269|0.479459|0.999999|0.739729|0.999989|0.000011|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.847505|0.391940|0.999999|0.695969|0.999987|0.000013|
|1|Radixor|PRIMARY_OUTPUT|0.999974|0.969675|1.000000|0.984838|0.999999|0.000001|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.921471|0.480558|0.999999|0.740279|0.999989|0.000011|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.940611|0.479343|0.999999|0.739671|0.999989|0.000011|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.890914|0.391451|0.999999|0.695725|0.999987|0.000013|
</details>
@@ -100,21 +108,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.975128|0.972893|0.970667|0.947216|0.972900|0.972899|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.758996|0.623613|0.529216|0.453080|0.653142|0.653138|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.769880|0.627374|0.529384|0.457061|0.659544|0.659540|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.687649|0.536000|0.439152|0.366120|0.576343|0.576338|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.972892|0.996085|0.993746|0.994914|0.994914|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.623608|0.990718|0.904385|0.945585|0.945585|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.627369|0.991872|0.904139|0.945975|0.945975|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.535994|0.988126|0.886473|0.934544|0.934544|
|1|Radixor|PRIMARY_OUTPUT|0.993763|0.984591|0.975587|0.969650|0.984708|0.984708|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.778598|0.631685|0.531413|0.461652|0.665448|0.665443|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.788800|0.635056|0.531468|0.465262|0.671472|0.671468|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.709787|0.543915|0.440884|0.373546|0.590550|0.590545|
</details>
@@ -122,62 +119,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|30552427|731279|971268|1641126083212|731279 / 1641126814491|971268 / 31523695|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|15153638|1922153|16370057|1641124892338|1922153 / 1641126814491|16370057 / 31523695|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|15114332|1544812|16409363|1641125269679|1544812 / 1641126814491|16409363 / 31523695|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|12355389|2223150|19168306|1641124591341|2223150 / 1641126814491|19168306 / 31523695|
|1|Radixor|PRIMARY_OUTPUT|30511413|804|954186|1599841738533|804 / 1599841739337|954186 / 31465599|
|2|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|15121052|1288634|16344547|1599840450703|1288634 / 1599841739337|16344547 / 31465599|
|3|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|15082807|952306|16382792|1599840787031|952306 / 1599841739337|16382792 / 31465599|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|12317229|1508153|19148370|1599840231184|1508153 / 1599841739337|19148370 / 31465599|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 1641126814491 (0.000000%)|0 / 31523695 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|31523695|0|0|1641126814491|0 / 1641126814491|0 / 31523695|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 1599841739337|0 / 31465599|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|1683575 / 1641126814491 (0.000103%)|0 / 31523695 (0.000000%)|0.959025|0.973991|0.974320|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
</div>
@@ -185,7 +160,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.949301|1.000000|0.999999|0.999999|0.999999|0.000001|
|1|Radixor|ALL_CANDIDATES|0.999926|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -193,15 +168,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.959025|0.973991|0.989432|0.949301|0.974321|0.974320|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999941|0.999963|0.999985|0.999926|0.999963|0.999963|
</details>
@@ -209,7 +176,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|31523695|1683575|0|1641125130916|1683575 / 1641126814491|0 / 31523695|
|1|Radixor|ALL_CANDIDATES|31465599|2327|0|1599841737010|2327 / 1599841739337|0 / 31465599|
</details>
@@ -219,22 +186,22 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|971268|731279|952296|57328|3.164291%|6|1876272|
|Radixor|954186|804|1523|34395|1.922815%|6|1826768|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988068|730145 / 1543589444152 (0.000047%)|735305 / 30813833 (2.386282%)|0.976268|0.976219|0.976218|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.738400|1513705 / 1543589444152 (0.000098%)|16121763 / 30813833 (52.319888%)|0.768117|0.624934|0.657464|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.738400|1513705 / 1543589444152 (0.000098%)|16121763 / 30813833 (52.319888%)|0.768117|0.624934|0.657464|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.694529|1806392 / 1543589444152 (0.000117%)|18825444 / 30813833 (61.094133%)|0.697056|0.537492|0.581469|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.988242|&lt;0.000001%|2.351587%|
|2|SNOWBALL FINNISH DIRECT|0.738344|0.000062%|52.331112%|
|3|SNOWBALL FINNISH LUCENE FILTER|0.738344|0.000062%|52.331112%|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|0.694308|0.000077%|61.138333%|
</div>
@@ -242,10 +209,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.976301|0.976137|1.000000|0.988068|0.999999|0.000001|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.906595|0.476801|0.999999|0.738400|0.999989|0.000011|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.906595|0.476801|0.999999|0.738400|0.999989|0.000011|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.869053|0.389059|0.999999|0.694529|0.999987|0.000013|
|1|Radixor|PRIMARY_OUTPUT|0.999973|0.976484|1.000000|0.988242|1.000000|0.000000|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.939951|0.476689|0.999999|0.738344|0.999989|0.000011|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.939951|0.476689|0.999999|0.738344|0.999989|0.000011|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.911893|0.388617|0.999999|0.694308|0.999987|0.000013|
</details>
@@ -253,21 +220,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.976268|0.976219|0.976170|0.953543|0.976219|0.976218|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.768117|0.624934|0.526744|0.454475|0.657469|0.657464|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.768117|0.624934|0.526744|0.454475|0.657469|0.657464|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.697056|0.537492|0.437372|0.367514|0.581474|0.581469|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.976218|0.996000|0.996069|0.996035|0.996035|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.624929|0.991732|0.902933|0.945252|0.945252|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.624929|0.991732|0.902933|0.945252|0.945252|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.537486|0.989268|0.885294|0.934397|0.934397|
|1|Radixor|PRIMARY_OUTPUT|0.995185|0.988089|0.981093|0.976459|0.988159|0.988159|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|0.786987|0.632573|0.528815|0.462601|0.669376|0.669372|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|0.786987|0.632573|0.528815|0.462601|0.669376|0.669372|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.718421|0.544981|0.438999|0.374553|0.595296|0.595291|
</details>
@@ -275,62 +231,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|30078528|730145|735305|1543588714007|730145 / 1543589444152|735305 / 30813833|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|14692070|1513705|16121763|1543587930447|1513705 / 1543589444152|16121763 / 30813833|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|14692070|1513705|16121763|1543587930447|1513705 / 1543589444152|16121763 / 30813833|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|11988389|1806392|18825444|1543587637760|1806392 / 1543589444152|18825444 / 30813833|
|1|Radixor|PRIMARY_OUTPUT|30037514|804|723369|1504706134249|804 / 1504706135053|723369 / 30760883|
|2|SNOWBALL FINNISH DIRECT|PRIMARY_OUTPUT|14663371|936765|16097512|1504705198288|936765 / 1504706135053|16097512 / 30760883|
|3|SNOWBALL FINNISH LUCENE FILTER|PRIMARY_OUTPUT|14663371|936765|16097512|1504705198288|936765 / 1504706135053|16097512 / 30760883|
|4|FINNISH LUCENE FINNISH LIGHT STEM FILTER|PRIMARY_OUTPUT|11954192|1155011|18806691|1504704980042|1155011 / 1504706135053|18806691 / 30760883|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 1543589444152 (0.000000%)|0 / 30813833 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|30813833|0|0|1543589444152|0 / 1543589444152|0 / 30813833|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 1504706135053|0 / 30760883|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|1653320 / 1543589444152 (0.000107%)|0 / 30813833 (0.000000%)|0.958843|0.973873|0.974205|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
</div>
@@ -338,7 +272,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.949077|1.000000|0.999999|0.999999|0.999999|0.000001|
|1|Radixor|ALL_CANDIDATES|0.999927|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -346,15 +280,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.958843|0.973873|0.989383|0.949077|0.974206|0.974205|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999942|0.999964|0.999985|0.999927|0.999964|0.999964|
</details>
@@ -362,7 +288,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|30813833|1653320|0|1543587790832|1653320 / 1543589444152|0 / 30813833|
|1|Radixor|ALL_CANDIDATES|30760883|2235|0|1504706132818|2235 / 1504706135053|0 / 30760883|
</details>
@@ -372,19 +298,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|735305|730145|923175|44331|2.523029%|6|1805864|
|Radixor|723369|804|1431|22060|1.271628%|6|1758300|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -393,16 +319,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `FI_FI`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

377
docs/benchmarks/languages/french.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `FR_FR` | 59,240 | 474,110 | 108,141 | 365,969 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `fr-fr-default` | `1.0.0` | `FR_FR` | 59,240 | 474,110 | 108,141 | 365,969 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **474,110**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **474,110**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 5,370 | 1.133% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 185,263 | 39.076% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 153,886 | 32.458% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 116,519 | 24.576% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 13,072 | 2.757% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 10,082 | 2.127% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 184,521 | 38.919% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 154,760 | 32.642% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 110,933 | 23.398% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 13,814 | 2.914% |
## Accuracy
@@ -37,18 +37,26 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 8.462% | 5.067% | 19.952% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene FrenchLightStemFilter | 6.377% | 3.965% | 14.540% | Light suffix stemmer; intentionally narrower than a dictionary-derived stemmer. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `frenchRadixor` | 47.033 | 4.146 | 128.5 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 1664.935 | 65.928 | 4549.4 | 35.399 | Benchmark-only French Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene FrenchMinimalStemFilter | `frenchLuceneFrenchMinimalStemFilter` | 19.234 | 2.098 | 52.6 | 0.409 | Minimal French suffix reducer; narrow baseline. |
| Lucene FrenchLightStemFilter | `frenchLuceneFrenchLightStemFilter` | 30.560 | 3.680 | 83.5 | 0.650 | Light French suffix stemmer. |
| Official Snowball direct | `snowballDirect[FRENCH]` | 111.057 | 8.172 | 303.5 | 2.361 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[FRENCH]` | 123.648 | 3.500 | 337.9 | 2.629 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `frenchRadixor` | 49.340 | 0.986 | 134.8 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 1781.070 | 43.544 | 4866.7 | 36.098 | Benchmark-only French Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene FrenchMinimalStemFilter | `frenchLuceneFrenchMinimalStemFilter` | 19.093 | 0.681 | 52.2 | 0.387 | Minimal French suffix reducer; narrow baseline. |
| Lucene FrenchLightStemFilter | `frenchLuceneFrenchLightStemFilter` | 29.553 | 0.465 | 80.8 | 0.599 | Light French suffix stemmer. |
| Official Snowball direct | `snowballDirect[FRENCH]` | 121.376 | 0.865 | 331.7 | 2.460 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[FRENCH]` | 126.574 | 4.671 | 345.9 | 2.565 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -62,32 +70,32 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `FR_FR` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `FR_FR` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/fr_fr/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `fr-fr-default`, loaded from classpath resource `org/egothor/stemmer/models/fr-fr-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.956992** among 6 deterministic stemmers. The runner-up is `SNOWBALL FRENCH DIRECT` at 0.845262, a difference of 0.111731. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.957224** among 6 deterministic stemmers. The runner-up is `SNOWBALL FRENCH DIRECT` at 0.845414, a difference of 0.111810. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.958627** among 6 deterministic stemmers. The runner-up is `SNOWBALL FRENCH DIRECT` at 0.848662, a difference of 0.109965. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.958856** among 6 deterministic stemmers. The runner-up is `SNOWBALL FRENCH DIRECT` at 0.848826, a difference of 0.110031. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.956992|318767 / 90396104830 (0.000353%)|469160 / 5454615 (8.601157%)|0.934603|0.926765|0.926851|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.845262|1654723 / 90396104830 (0.001831%)|1687975 / 5454615 (30.945814%)|0.693926|0.692653|0.692638|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.844999|1661388 / 90396104830 (0.001838%)|1690838 / 5454615 (30.998301%)|0.693010|0.691885|0.691869|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.813742|776728 / 90396104830 (0.000859%)|2031881 / 5454615 (37.250677%)|0.769069|0.709075|0.715131|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.518587|276403 / 90396104830 (0.000306%)|5251833 / 5454615 (96.282377%)|0.137547|0.068348|0.125415|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.516830|160438 / 90396104830 (0.000177%)|5271003 / 5454615 (96.633823%)|0.134400|0.063329|0.134021|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.958627|&lt;0.000001%|8.274665%|
|2|SNOWBALL FRENCH DIRECT|0.848662|0.001338%|30.266309%|
|3|SNOWBALL FRENCH LUCENE FILTER|0.848404|0.001345%|30.317815%|
|4|HUNSPELL FRENCH LUCENE FILTER|0.816824|0.000540%|36.634583%|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|0.518478|0.000187%|96.304159%|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|0.516784|0.000083%|96.643216%|
</div>
@@ -95,12 +103,12 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.939903|0.913988|0.999996|0.956992|0.999991|0.000009|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.694777|0.690542|0.999982|0.845262|0.999963|0.000037|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.693763|0.690017|0.999982|0.844999|0.999963|0.000037|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.815041|0.627493|0.999991|0.813742|0.999969|0.000031|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.423181|0.037176|0.999997|0.518587|0.999939|0.000061|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.533678|0.033662|0.999998|0.516830|0.999940|0.000060|
|1|Radixor|PRIMARY_OUTPUT|0.999994|0.917253|1.000000|0.958627|0.999995|0.000005|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.774195|0.697337|0.999987|0.848662|0.999967|0.000033|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.773169|0.696822|0.999987|0.848404|0.999967|0.000033|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.885315|0.633654|0.999995|0.816824|0.999970|0.000030|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.565022|0.036958|0.999998|0.518478|0.999935|0.000065|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.726387|0.033568|0.999999|0.516784|0.999936|0.000064|
</details>
@@ -108,25 +116,12 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.934603|0.926765|0.919056|0.863524|0.926855|0.926851|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.693926|0.692653|0.691385|0.529816|0.692656|0.692638|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.693010|0.691885|0.690763|0.528917|0.691887|0.691869|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.769069|0.709075|0.657765|0.549277|0.715145|0.715131|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.137547|0.068348|0.045472|0.035383|0.125428|0.125415|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.134400|0.063329|0.041424|0.032700|0.134032|0.134021|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.926760|0.988772|0.985214|0.986990|0.986990|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.692635|0.959459|0.944948|0.952148|0.952148|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.691866|0.958698|0.944715|0.951655|0.951655|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.709060|0.978337|0.913706|0.944918|0.944918|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.068339|0.974110|0.812376|0.885922|0.885922|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.063322|0.984019|0.810979|0.889158|0.889158|
|1|Radixor|PRIMARY_OUTPUT|0.982273|0.956838|0.932688|0.917248|0.957731|0.957728|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.757497|0.733759|0.711463|0.579478|0.734761|0.734745|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.756590|0.733013|0.710861|0.578548|0.734003|0.733987|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.820168|0.738637|0.671850|0.585587|0.748988|0.748975|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.146469|0.069379|0.045455|0.035936|0.144507|0.144495|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.141655|0.064170|0.041481|0.033149|0.156151|0.156143|
</details>
@@ -134,70 +129,45 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|4985455|318767|469160|90395786063|318767 / 90396104830|469160 / 5454615|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|3766640|1654723|1687975|90394450107|1654723 / 90396104830|1687975 / 5454615|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3763777|1661388|1690838|90394443442|1661388 / 90396104830|1690838 / 5454615|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3422734|776728|2031881|90395328102|776728 / 90396104830|2031881 / 5454615|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|202782|276403|5251833|90395828427|276403 / 90396104830|5251833 / 5454615|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|183612|160438|5271003|90395944392|160438 / 90396104830|5271003 / 5454615|
|1|Radixor|PRIMARY_OUTPUT|4925833|29|444366|81606871827|29 / 81606871856|444366 / 5370199|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|3744838|1092238|1625361|81605779618|1092238 / 81606871856|1625361 / 5370199|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3742072|1097843|1628127|81605774013|1097843 / 81606871856|1628127 / 5370199|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3402849|440809|1967350|81606431047|440809 / 81606871856|1967350 / 5370199|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|198474|152794|5171725|81606719062|152794 / 81606871856|5171725 / 5370199|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|180266|67902|5189933|81606803954|67902 / 81606871856|5189933 / 5370199|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999979|12 / 90396104830 (0.000000%)|232 / 5454615 (0.004253%)|0.999990|0.999978|0.999978|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|0.830964|745831 / 90396104830 (0.000825%)|1844003 / 5454615 (33.806291%)|0.789019|0.736029|0.740670|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.004320%|
|HUNSPELL FRENCH LUCENE FILTER|0.000539%|33.189869%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999998|0.999957|1.000000|0.999979|1.000000|0.000000|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|0.828798|0.661937|0.999992|0.830964|0.999971|0.000029|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999990|0.999978|0.999966|0.999955|0.999978|0.999978|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|0.789019|0.736029|0.689709|0.582315|0.740684|0.740670|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|5454383|12|232|90396104818|12 / 90396104830|232 / 5454615|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|3610612|745831|1844003|90395358999|745831 / 90396104830|1844003 / 5454615|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 81606871856|232 / 5370199|
|HUNSPELL FRENCH LUCENE FILTER|439665 / 81606871856|1782362 / 5370199|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999973|1056255 / 90396104830 (0.001168%)|232 / 5454615 (0.004253%)|0.865853|0.911704|0.915270|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.830963|1043199 / 90396104830 (0.001154%)|1844003 / 5454615 (33.806291%)|0.750028|0.714377|0.716613|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.999978|0.000003%|0.004320%|
|2|HUNSPELL FRENCH LUCENE FILTER|0.834048|0.000614%|33.189869%|
</div>
@@ -205,8 +175,8 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.837765|0.999957|0.999988|0.999973|0.999988|0.000012|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.775840|0.661937|0.999988|0.830963|0.999968|0.000032|
|1|Radixor|ALL_CANDIDATES|0.999571|0.999957|1.000000|0.999978|1.000000|0.000000|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.877537|0.668101|0.999994|0.834048|0.999972|0.000028|
</details>
@@ -214,17 +184,8 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.865853|0.911704|0.962682|0.837735|0.915275|0.915270|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.750028|0.714377|0.681961|0.555666|0.716629|0.716613|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999648|0.999764|0.999880|0.999528|0.999764|0.999764|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.825765|0.758630|0.701590|0.611123|0.765691|0.765678|
</details>
@@ -232,8 +193,8 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|5454383|1056255|232|90395048575|1056255 / 90396104830|232 / 5454615|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|3610612|1043199|1844003|90395061631|1043199 / 90396104830|1844003 / 5454615|
|1|Radixor|ALL_CANDIDATES|5369967|2303|232|81606869553|2303 / 81606871856|232 / 5370199|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|3587837|500695|1782362|81606371161|500695 / 81606871856|1782362 / 5370199|
</details>
@@ -243,25 +204,25 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|468928|318755|737488|43040|10.122057%|56|477024|
|HUNSPELL FRENCH LUCENE FILTER|187878|30897|266471|13511|3.177489%|4|439015|
|Radixor|444134|29|2274|21844|5.406783%|56|427440|
|HUNSPELL FRENCH LUCENE FILTER|184988|1144|59886|8230|2.037073%|4|412364|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.957224|315266 / 88712126506 (0.000355%)|465436 / 5440559 (8.554930%)|0.935099|0.927248|0.927334|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.845414|1646111 / 88712126506 (0.001856%)|1681970 / 5440559 (30.915389%)|0.694508|0.693130|0.693115|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.845163|1641925 / 88712126506 (0.001851%)|1684703 / 5440559 (30.965623%)|0.694714|0.693068|0.693055|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.813617|763305 / 88712126506 (0.000860%)|2028011 / 5440559 (37.275784%)|0.770537|0.709734|0.715938|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.518442|262689 / 88712126506 (0.000296%)|5239869 / 5440559 (96.311225%)|0.137571|0.067985|0.126383|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.516697|147476 / 88712126506 (0.000166%)|5258873 / 5440559 (96.660527%)|0.134439|0.062979|0.135757|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.958856|&lt;0.000001%|8.228703%|
|2|SNOWBALL FRENCH DIRECT|0.848826|0.001356%|30.233460%|
|3|SNOWBALL FRENCH LUCENE FILTER|0.848580|0.001353%|30.282729%|
|4|HUNSPELL FRENCH LUCENE FILTER|0.816702|0.000540%|36.658999%|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|0.518338|0.000181%|96.332173%|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|0.516654|0.000076%|96.669051%|
</div>
@@ -269,12 +230,12 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.940408|0.914451|0.999996|0.957224|0.999991|0.000009|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.695430|0.690846|0.999981|0.845414|0.999962|0.000038|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.695815|0.690344|0.999981|0.845163|0.999963|0.000037|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.817210|0.627242|0.999991|0.813617|0.999969|0.000031|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.433101|0.036888|0.999997|0.518442|0.999938|0.000062|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.551965|0.033395|0.999998|0.516697|0.999939|0.000061|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.917713|1.000000|0.958856|0.999995|0.000005|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.774357|0.697665|0.999986|0.848826|0.999966|0.000034|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.774620|0.697173|0.999986|0.848580|0.999966|0.000034|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.886736|0.633410|0.999995|0.816702|0.999970|0.000030|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.575115|0.036678|0.999998|0.518338|0.999934|0.000066|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.746071|0.033309|0.999999|0.516654|0.999935|0.000065|
</details>
@@ -282,25 +243,12 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.935099|0.927248|0.919527|0.864363|0.927338|0.927334|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.694508|0.693130|0.691758|0.530374|0.693134|0.693115|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.694714|0.693068|0.691431|0.530302|0.693074|0.693055|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.770537|0.709734|0.657826|0.550068|0.715953|0.715938|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.137571|0.067985|0.045148|0.035189|0.126397|0.126383|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.134439|0.062979|0.041121|0.032513|0.135767|0.135757|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.927243|0.988916|0.985550|0.987230|0.987230|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.693112|0.959521|0.944537|0.951970|0.951970|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.693050|0.959566|0.944385|0.951915|0.951915|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.709719|0.979328|0.913162|0.945088|0.945088|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.067976|0.975086|0.811144|0.885591|0.885591|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.062973|0.985086|0.809774|0.888868|0.888868|
|1|Radixor|PRIMARY_OUTPUT|0.982383|0.957091|0.933069|0.917713|0.957973|0.957971|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|0.757699|0.734013|0.711764|0.579795|0.735012|0.734995|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.757784|0.733859|0.711398|0.579603|0.734877|0.734860|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|0.821061|0.738965|0.671794|0.585999|0.749445|0.749431|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.146117|0.068959|0.045128|0.035711|0.145238|0.145227|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.141311|0.063772|0.041177|0.032936|0.157643|0.157634|
</details>
@@ -308,70 +256,45 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|4975123|315266|465436|88711811240|315266 / 88712126506|465436 / 5440559|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|3758589|1646111|1681970|88710480395|1646111 / 88712126506|1681970 / 5440559|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3755856|1641925|1684703|88710484581|1641925 / 88712126506|1684703 / 5440559|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3412548|763305|2028011|88711363201|763305 / 88712126506|2028011 / 5440559|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|200690|262689|5239869|88711863817|262689 / 88712126506|5239869 / 5440559|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|181686|147476|5258873|88711979030|147476 / 88712126506|5258873 / 5440559|
|1|Radixor|PRIMARY_OUTPUT|4915501|1|440750|80279496864|1 / 80279496865|440750 / 5356251|
|2|SNOWBALL FRENCH DIRECT|PRIMARY_OUTPUT|3736871|1088903|1619380|80278407962|1088903 / 80279496865|1619380 / 5356251|
|3|SNOWBALL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3734232|1086494|1622019|80278410371|1086494 / 80279496865|1622019 / 5356251|
|4|HUNSPELL FRENCH LUCENE FILTER|PRIMARY_OUTPUT|3392703|433354|1963548|80279063511|433354 / 80279496865|1963548 / 5356251|
|5|FRENCH LUCENE FRENCH LIGHT STEM FILTER|PRIMARY_OUTPUT|196458|145140|5159793|80279351725|145140 / 80279496865|5159793 / 5356251|
|6|FRENCH LUCENE FRENCH MINIMAL STEM FILTER|PRIMARY_OUTPUT|178414|60724|5177837|80279436141|60724 / 80279496865|5177837 / 5356251|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 88712126506 (0.000000%)|0 / 5440559 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|0.830852|733584 / 88712126506 (0.000827%)|1840476 / 5440559 (33.828803%)|0.790351|0.736648|0.741404|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
|HUNSPELL FRENCH LUCENE FILTER|0.000539%|33.211718%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|0.830724|0.661712|0.999992|0.830852|0.999971|0.000029|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|0.790351|0.736648|0.689779|0.583090|0.741418|0.741404|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|5440559|0|0|88712126506|0 / 88712126506|0 / 5440559|
|2|HUNSPELL FRENCH LUCENE FILTER|ANY_CANDIDATE|3600083|733584|1840476|88711392922|733584 / 88712126506|1840476 / 5440559|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 80279496865|0 / 5356251|
|HUNSPELL FRENCH LUCENE FILTER|432307 / 80279496865|1778903 / 5356251|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999995|938985 / 88712126506 (0.001058%)|0 / 5440559 (0.000000%)|0.878679|0.920560|0.923474|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.830850|1027635 / 88712126506 (0.001158%)|1840476 / 5440559 (33.828803%)|0.751538|0.715134|0.717460|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
|2|HUNSPELL FRENCH LUCENE FILTER|0.833938|0.000614%|33.211718%|
</div>
@@ -379,8 +302,8 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.852813|1.000000|0.999989|0.999995|0.999989|0.000011|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.777939|0.661712|0.999988|0.830850|0.999968|0.000032|
|1|Radixor|ALL_CANDIDATES|0.999986|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.878983|0.667883|0.999994|0.833938|0.999972|0.000028|
</details>
@@ -388,17 +311,8 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.878679|0.920560|0.966634|0.852813|0.923479|0.923474|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.751538|0.715134|0.682093|0.556582|0.717476|0.717460|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999989|0.999993|0.999997|0.999986|0.999993|0.999993|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|0.826722|0.759029|0.701582|0.611641|0.766197|0.766184|
</details>
@@ -406,8 +320,8 @@ This mode contains **10 result rows**, **6 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|5440559|938985|0|88711187521|938985 / 88712126506|0 / 5440559|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|3600083|1027635|1840476|88711098871|1027635 / 88712126506|1840476 / 5440559|
|1|Radixor|ALL_CANDIDATES|5356251|75|0|80279496790|75 / 80279496865|0 / 5356251|
|2|HUNSPELL FRENCH LUCENE FILTER|ALL_CANDIDATES|3577348|492522|1778903|80279004343|492522 / 80279496865|1778903 / 5356251|
</details>
@@ -417,20 +331,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|465436|315266|623719|41130|9.764239%|56|468574|
|HUNSPELL FRENCH LUCENE FILTER|187535|29721|264330|13437|3.189936%|4|434961|
|Radixor|440750|1|74|20611|5.143594%|56|422336|
|HUNSPELL FRENCH LUCENE FILTER|184645|1047|59168|8194|2.044860%|4|409028|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -439,16 +353,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `FR_FR`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

417
docs/benchmarks/languages/german.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `DE_DE` | 39,315 | 213,440 | 73,799 | 139,641 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `de-de-default` | `1.0.0` | `DE_DE` | 54,092 | 333,036 | 90,535 | 242,501 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **213,440**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **333,036**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 3,627 | 1.699% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 48,605 | 22.772% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 80,443 | 37.689% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 75,717 | 35.475% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 5,048 | 2.365% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 12,107 | 3.635% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 81,805 | 24.563% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 142,376 | 42.751% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 88,820 | 26.670% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 7,928 | 2.381% |
## Accuracy
@@ -39,20 +39,28 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 30.481% | 29.027% | 34.376% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene GermanStemFilter | 21.559% | 19.312% | 27.576% | German Lucene stemming TokenFilter; broader than minimal/light variants. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `germanRadixor` | 41.166 | 2.396 | 294.8 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| CISTEM | `germanCistem` | 248.392 | 12.294 | 1778.8 | 6.034 | Benchmark-only CISTEM implementation. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 281.322 | 3.411 | 2014.6 | 6.834 | Benchmark-only German Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene GermanMinimalStemFilter | `germanLuceneGermanMinimalStemFilter` | 23.562 | 0.969 | 168.7 | 0.572 | Minimal German suffix reduction; narrow baseline. |
| Lucene GermanLightStemFilter | `germanLuceneGermanLightStemFilter` | 24.410 | 1.034 | 174.8 | 0.593 | Light German suffix stemmer; narrower than a dictionary stemmer. |
| Lucene GermanStemFilter | `germanLuceneGermanStemFilter` | 71.039 | 4.443 | 508.7 | 1.726 | Older German stemming TokenFilter with normalization requirements. |
| Lucene SnowballFilter | `luceneSnowballFilter[GERMAN]` | 105.771 | 9.617 | 757.4 | 2.569 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Official Snowball direct | `snowballDirect[GERMAN]` | 100.688 | 9.018 | 721.0 | 2.446 | Official Snowball generated Java stemmer; direct API. |
| Radixor | `germanRadixor` | 40.571 | 1.647 | 167.3 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| CISTEM | `germanCistem` | 305.166 | 4.590 | 1258.4 | 7.522 | Benchmark-only CISTEM implementation. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 291.791 | 21.769 | 1203.3 | 7.192 | Benchmark-only German Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene GermanMinimalStemFilter | `germanLuceneGermanMinimalStemFilter` | 23.903 | 0.208 | 98.6 | 0.589 | Minimal German suffix reduction; narrow baseline. |
| Lucene GermanLightStemFilter | `germanLuceneGermanLightStemFilter` | 24.695 | 0.322 | 101.8 | 0.609 | Light German suffix stemmer; narrower than a dictionary stemmer. |
| Lucene GermanStemFilter | `germanLuceneGermanStemFilter` | 72.140 | 1.544 | 297.5 | 1.778 | Older German stemming TokenFilter with normalization requirements. |
| Lucene SnowballFilter | `luceneSnowballFilter[GERMAN]` | 110.086 | 2.315 | 454.0 | 2.713 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Official Snowball direct | `snowballDirect[GERMAN]` | 100.122 | 2.623 | 412.9 | 2.468 | Official Snowball generated Java stemmer; direct API. |
## Interpretation Notes
@@ -66,34 +74,34 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `DE_DE` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `DE_DE` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/de_de/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `de-de-default`, loaded from classpath resource `org/egothor/stemmer/models/de-de-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.907901** among 8 deterministic stemmers. The runner-up is `GERMAN CISTEM` at 0.880770, a difference of 0.027131. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.966157** among 8 deterministic stemmers. The runner-up is `GERMAN CISTEM` at 0.915288, a difference of 0.050869. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.910445** among 8 deterministic stemmers. The runner-up is `GERMAN CISTEM` at 0.878527, a difference of 0.031918. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.966959** among 8 deterministic stemmers. The runner-up is `GERMAN CISTEM` at 0.914727, a difference of 0.052232. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.907901|98192 / 44095245979 (0.000223%)|254903 / 1383872 (18.419550%)|0.897073|0.864768|0.866326|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.880770|477122 / 44095245979 (0.001082%)|329983 / 1383872 (23.844908%)|0.701852|0.723109|0.724023|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.778614|190680 / 44095245979 (0.000432%)|612734 / 1383872 (44.276783%)|0.737064|0.657494|0.668394|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.771357|295701 / 44095245979 (0.000671%)|632816 / 1383872 (45.727929%)|0.674089|0.617993|0.624014|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.756258|205740 / 44095245979 (0.000467%)|674609 / 1383872 (48.747933%)|0.703092|0.617052|0.630292|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.723772|331871 / 44095245979 (0.000753%)|764518 / 1383872 (55.244849%)|0.596821|0.530474|0.539809|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.641579|203883 / 44095245979 (0.000462%)|992010 / 1383872 (71.683653%)|0.520145|0.395897|0.431563|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.598139|110840 / 44095245979 (0.000251%)|1112246 / 1383872 (80.372029%)|0.466113|0.307558|0.373350|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.910445|0.000002%|17.910967%|
|2|GERMAN CISTEM|0.878527|0.000674%|24.293900%|
|3|SNOWBALL GERMAN DIRECT|0.776006|0.000171%|44.798684%|
|4|SNOWBALL GERMAN LUCENE FILTER|0.769071|0.000371%|46.185528%|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|0.753833|0.000191%|49.233299%|
|6|GERMAN LUCENE GERMAN STEM FILTER|0.720992|0.000443%|55.801084%|
|7|HUNSPELL GERMAN LUCENE FILTER|0.640308|0.000290%|71.938102%|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|0.595748|0.000088%|80.850384%|
</div>
@@ -101,14 +109,14 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.919984|0.815804|0.999998|0.907901|0.999992|0.000008|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.688361|0.761551|0.999989|0.880770|0.999982|0.000018|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.801750|0.557232|0.999996|0.778614|0.999982|0.000018|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.717508|0.542721|0.999993|0.771357|0.999979|0.000021|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.775148|0.512521|0.999995|0.756258|0.999980|0.000020|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.651112|0.447552|0.999992|0.723772|0.999975|0.000025|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.657768|0.283163|0.999995|0.641579|0.999973|0.000027|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.710196|0.196280|0.999997|0.598139|0.999972|0.000028|
|1|Radixor|PRIMARY_OUTPUT|0.999400|0.820890|1.000000|0.910445|0.999994|0.000006|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.797231|0.757061|0.999993|0.878527|0.999985|0.000015|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.918570|0.552013|0.999998|0.776006|0.999983|0.000017|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.835220|0.538145|0.999996|0.769071|0.999980|0.000020|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.902792|0.507667|0.999998|0.753833|0.999981|0.000019|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.777304|0.441989|0.999996|0.720992|0.999976|0.000024|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.771720|0.280619|0.999997|0.640308|0.999972|0.000028|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.883845|0.191496|0.999999|0.595748|0.999971|0.000029|
</details>
@@ -116,29 +124,14 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.897073|0.864768|0.834709|0.761755|0.866330|0.866326|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.701852|0.723109|0.745694|0.566304|0.724032|0.724023|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.737064|0.657494|0.593429|0.489751|0.668402|0.668394|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.674089|0.617993|0.570517|0.447171|0.624024|0.624014|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.703092|0.617052|0.549774|0.446186|0.630301|0.630292|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.596821|0.530474|0.477402|0.360983|0.539820|0.539809|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.520145|0.395897|0.319562|0.246803|0.431574|0.431563|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.466113|0.307558|0.229493|0.181725|0.373359|0.373350|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.864764|0.989946|0.975085|0.982460|0.982460|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.723100|0.974048|0.975147|0.974597|0.974597|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.657485|0.983725|0.949324|0.966218|0.966218|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.617983|0.975845|0.942925|0.959102|0.959102|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.617043|0.980753|0.936533|0.958133|0.958133|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.530462|0.975550|0.942890|0.958942|0.958942|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.395885|0.980463|0.886873|0.931322|0.931322|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.307549|0.983615|0.896264|0.937910|0.937910|
|1|Radixor|PRIMARY_OUTPUT|0.957746|0.901392|0.851302|0.820486|0.905758|0.905755|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.788860|0.776627|0.764768|0.634824|0.776886|0.776879|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.810879|0.689608|0.599891|0.526260|0.712083|0.712076|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.752175|0.654552|0.579359|0.486494|0.670425|0.670416|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.781189|0.649884|0.556368|0.481355|0.676991|0.676984|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.674901|0.563540|0.483723|0.392311|0.586140|0.586130|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.571639|0.411577|0.321543|0.259110|0.465359|0.465349|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.512941|0.314789|0.227071|0.186795|0.411404|0.411396|
</details>
@@ -146,72 +139,47 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|1128969|98192|254903|44095147787|98192 / 44095245979|254903 / 1383872|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|1053889|477122|329983|44094768857|477122 / 44095245979|329983 / 1383872|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|771138|190680|612734|44095055299|190680 / 44095245979|612734 / 1383872|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|751056|295701|632816|44094950278|295701 / 44095245979|632816 / 1383872|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|709263|205740|674609|44095040239|205740 / 44095245979|674609 / 1383872|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|619354|331871|764518|44094914108|331871 / 44095245979|764518 / 1383872|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|391862|203883|992010|44095042096|203883 / 44095245979|992010 / 1383872|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|271626|110840|1112246|44095135139|110840 / 44095245979|1112246 / 1383872|
|1|Radixor|PRIMARY_OUTPUT|1103976|663|240876|38436733230|663 / 38436733893|240876 / 1344852|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|1018135|258954|326717|38436474939|258954 / 38436733893|326717 / 1344852|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|742376|65811|602476|38436668082|65811 / 38436733893|602476 / 1344852|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|723725|142783|621127|38436591110|142783 / 38436733893|621127 / 1344852|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|682737|73514|662115|38436660379|73514 / 38436733893|662115 / 1344852|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|594410|170297|750442|38436563596|170297 / 38436733893|750442 / 1344852|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|377391|111635|967461|38436622258|111635 / 38436733893|967461 / 1344852|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|257534|33845|1087318|38436700048|33845 / 38436733893|1087318 / 1344852|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.959835|1375 / 44095245979 (0.000003%)|111167 / 1383872 (8.033041%)|0.981996|0.957658|0.958475|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|0.647474|158403 / 44095245979 (0.000359%)|975697 / 1383872 (70.504859%)|0.559116|0.418544|0.460956|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000001%|8.261653%|
|HUNSPELL GERMAN LUCENE FILTER|0.000216%|70.811435%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.998921|0.919670|1.000000|0.959835|0.999997|0.000003|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|0.720422|0.294951|0.999996|0.647474|0.999974|0.000026|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.981996|0.957658|0.934498|0.918757|0.958476|0.958475|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|0.559116|0.418544|0.334456|0.264658|0.460966|0.460956|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1272705|1375|111167|44095244604|1375 / 44095245979|111167 / 1383872|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|408175|158403|975697|44095087576|158403 / 44095245979|975697 / 1383872|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|502 / 38436733893|111107 / 1344852|
|HUNSPELL GERMAN LUCENE FILTER|83073 / 38436733893|952309 / 1344852|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.959832|244817 / 44095245979 (0.000555%)|111167 / 1383872 (8.033041%)|0.853711|0.877306|0.878234|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.647473|242551 / 44095245979 (0.000550%)|975697 / 1383872 (70.504859%)|0.511911|0.401234|0.430118|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.958692|0.000018%|8.261653%|
|2|HUNSPELL GERMAN LUCENE FILTER|0.645941|0.000354%|70.811435%|
</div>
@@ -219,8 +187,8 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.838673|0.919670|0.999994|0.959832|0.999992|0.000008|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.627261|0.294951|0.999994|0.647473|0.999972|0.000028|
|1|Radixor|ALL_CANDIDATES|0.994469|0.917383|1.000000|0.958692|0.999997|0.000003|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.742744|0.291886|0.999996|0.645941|0.999972|0.000028|
</details>
@@ -228,17 +196,8 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.853711|0.877306|0.902242|0.781429|0.878238|0.878234|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.511911|0.401234|0.329907|0.250965|0.430130|0.430118|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.978033|0.954372|0.931829|0.912726|0.955149|0.955147|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.567444|0.419080|0.332218|0.265086|0.465614|0.465603|
</details>
@@ -246,8 +205,8 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1272705|244817|111167|44095001162|244817 / 44095245979|111167 / 1383872|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|408175|242551|975697|44095003428|242551 / 44095245979|975697 / 1383872|
|1|Radixor|ALL_CANDIDATES|1233745|6862|111107|38436727031|6862 / 38436733893|111107 / 1344852|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|392543|135961|952309|38436597932|135961 / 38436733893|952309 / 1344852|
</details>
@@ -257,27 +216,27 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|143736|96817|146625|48574|16.356314%|8|361016|
|HUNSPELL GERMAN LUCENE FILTER|16313|45480|38668|7891|2.657135%|3|305052|
|Radixor|129769|161|6199|29035|10.471893%|8|313927|
|HUNSPELL GERMAN LUCENE FILTER|15152|28562|24326|6482|2.337827%|3|283881|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.966157|47898 / 11263756342 (0.000425%)|59114 / 873411 (6.768177%)|0.941996|0.938343|0.938358|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.915288|156784 / 11263756342 (0.001392%)|147964 / 873411 (16.940936%)|0.823934|0.826418|0.826415|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.795926|87697 / 11263756342 (0.000779%)|356475 / 873411 (40.814118%)|0.785153|0.699487|0.711329|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.775641|77653 / 11263756342 (0.000689%)|391910 / 873411 (44.871200%)|0.774111|0.672222|0.688986|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.769953|55477 / 11263756342 (0.000493%)|401846 / 873411 (46.008809%)|0.790797|0.673446|0.695023|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.716810|78723 / 11263756342 (0.000699%)|494677 / 873411 (56.637368%)|0.700519|0.569153|0.599149|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.659196|84679 / 11263756342 (0.000752%)|595318 / 873411 (68.160122%)|0.598178|0.449922|0.494019|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.575691|21214 / 11263756342 (0.000188%)|741190 / 873411 (84.861537%)|0.444545|0.257528|0.361168|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.966959|0.000001%|6.608210%|
|2|GERMAN CISTEM|0.914727|0.000812%|17.053716%|
|3|SNOWBALL GERMAN DIRECT|0.794994|0.000391%|41.000819%|
|4|SNOWBALL GERMAN LUCENE FILTER|0.774716|0.000325%|45.056540%|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|0.768968|0.000130%|46.206331%|
|6|GERMAN LUCENE GERMAN STEM FILTER|0.716147|0.000358%|56.770194%|
|7|HUNSPELL GERMAN LUCENE FILTER|0.659574|0.000556%|68.084626%|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|0.574999|0.000045%|85.000064%|
</div>
@@ -285,14 +244,14 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.944446|0.932318|0.999996|0.966157|0.999991|0.000009|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.822287|0.830591|0.999986|0.915288|0.999973|0.000027|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.854958|0.591859|0.999992|0.795926|0.999961|0.000039|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.861124|0.551288|0.999993|0.775641|0.999958|0.000042|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.894739|0.539912|0.999995|0.769953|0.999959|0.000041|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.827912|0.433626|0.999993|0.716810|0.999949|0.000051|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.766578|0.318399|0.999992|0.659196|0.999940|0.000060|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.861739|0.151385|0.999998|0.575691|0.999932|0.000068|
|1|Radixor|PRIMARY_OUTPUT|0.999900|0.933918|1.000000|0.966959|0.999995|0.000005|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.892172|0.829463|0.999992|0.914727|0.999978|0.000022|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.924303|0.589992|0.999996|0.794994|0.999963|0.000037|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.931871|0.549435|0.999997|0.774716|0.999960|0.000040|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.971001|0.537937|0.999999|0.768968|0.999961|0.000039|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.907196|0.432298|0.999996|0.716147|0.999950|0.000050|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.823043|0.319154|0.999994|0.659574|0.999939|0.000061|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.964480|0.149999|1.000000|0.574999|0.999931|0.000069|
</details>
@@ -300,29 +259,14 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.941996|0.938343|0.934719|0.883848|0.938363|0.938358|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.823934|0.826418|0.828917|0.704184|0.826428|0.826415|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.785153|0.699487|0.630675|0.537854|0.711347|0.711329|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.774111|0.672222|0.594035|0.506276|0.689005|0.688986|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.790797|0.673446|0.586424|0.507666|0.695040|0.695023|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.700519|0.569153|0.479277|0.397774|0.599170|0.599149|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.598178|0.449922|0.360559|0.290258|0.494042|0.494019|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.444545|0.257528|0.181270|0.147795|0.361184|0.361168|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.938338|0.994062|0.990664|0.992360|0.992360|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.826404|0.985936|0.973570|0.979714|0.979714|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.699468|0.988418|0.932452|0.959619|0.959619|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.672202|0.989021|0.919542|0.953017|0.953017|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.673427|0.991320|0.915070|0.951670|0.951670|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.569130|0.988584|0.918718|0.952371|0.952371|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.449897|0.988041|0.865581|0.922766|0.922766|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.257511|0.992643|0.854403|0.918349|0.918349|
|1|Radixor|PRIMARY_OUTPUT|0.985968|0.965783|0.946408|0.933831|0.966346|0.966343|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|0.878883|0.859676|0.841289|0.753887|0.860246|0.860236|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|0.830217|0.720244|0.635999|0.562799|0.738466|0.738450|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.817997|0.691285|0.598564|0.528217|0.715543|0.715527|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.836342|0.692324|0.590620|0.529431|0.722729|0.722714|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|0.743781|0.585563|0.482850|0.413990|0.626242|0.626223|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|0.625524|0.459951|0.363685|0.298660|0.512520|0.512499|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.462363|0.259621|0.180482|0.149175|0.380357|0.380343|
</details>
@@ -330,72 +274,47 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|814297|47898|59114|11263708444|47898 / 11263756342|59114 / 873411|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|725447|156784|147964|11263599558|156784 / 11263756342|147964 / 873411|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|516936|87697|356475|11263668645|87697 / 11263756342|356475 / 873411|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|481501|77653|391910|11263678689|77653 / 11263756342|391910 / 873411|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|471565|55477|401846|11263700865|55477 / 11263756342|401846 / 873411|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|378734|78723|494677|11263677619|78723 / 11263756342|494677 / 873411|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|278093|84679|595318|11263671663|84679 / 11263756342|595318 / 873411|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|132221|21214|741190|11263735128|21214 / 11263756342|741190 / 873411|
|1|Radixor|PRIMARY_OUTPUT|801691|80|56726|10594963454|80 / 10594963534|56726 / 858417|
|2|GERMAN CISTEM|PRIMARY_OUTPUT|712025|86055|146392|10594877479|86055 / 10594963534|146392 / 858417|
|3|SNOWBALL GERMAN DIRECT|PRIMARY_OUTPUT|506459|41477|351958|10594922057|41477 / 10594963534|351958 / 858417|
|4|SNOWBALL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|471644|34482|386773|10594929052|34482 / 10594963534|386773 / 858417|
|5|GERMAN LUCENE GERMAN LIGHT STEM FILTER|PRIMARY_OUTPUT|461774|13791|396643|10594949743|13791 / 10594963534|396643 / 858417|
|6|GERMAN LUCENE GERMAN STEM FILTER|PRIMARY_OUTPUT|371092|37962|487325|10594925572|37962 / 10594963534|487325 / 858417|
|7|HUNSPELL GERMAN LUCENE FILTER|PRIMARY_OUTPUT|273967|58904|584450|10594904630|58904 / 10594963534|584450 / 858417|
|8|GERMAN LUCENE GERMAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|128762|4742|729655|10594958792|4742 / 10594963534|729655 / 858417|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 11263756342 (0.000000%)|0 / 873411 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|0.665363|60996 / 11263756342 (0.000542%)|584547 / 873411 (66.926911%)|0.635466|0.472281|0.522540|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
|HUNSPELL GERMAN LUCENE FILTER|0.000383%|66.866802%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|0.825656|0.330731|0.999995|0.665363|0.999943|0.000057|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|0.635466|0.472281|0.375782|0.309142|0.522561|0.522540|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|873411|0|0|11263756342|0 / 11263756342|0 / 873411|
|2|HUNSPELL GERMAN LUCENE FILTER|ANY_CANDIDATE|288864|60996|584547|11263695346|60996 / 11263756342|584547 / 873411|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 10594963534|0 / 858417|
|HUNSPELL GERMAN LUCENE FILTER|40608 / 10594963534|573996 / 858417|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999996|97544 / 11263756342 (0.000866%)|0 / 873411 (0.000000%)|0.917983|0.947112|0.948436|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.665361|96545 / 11263756342 (0.000857%)|584547 / 873411 (66.926911%)|0.598050|0.458944|0.497855|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000014%|0.000000%|
|2|HUNSPELL GERMAN LUCENE FILTER|0.665663|0.000629%|66.866802%|
</div>
@@ -403,8 +322,8 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.899538|1.000000|0.999991|0.999996|0.999991|0.000009|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.749500|0.330731|0.999991|0.665361|0.999940|0.000060|
|1|Radixor|ALL_CANDIDATES|0.998267|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.810178|0.331332|0.999994|0.665663|0.999940|0.000060|
</details>
@@ -412,17 +331,8 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.917983|0.947112|0.978152|0.899538|0.948440|0.948436|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.598050|0.458944|0.372338|0.297811|0.497878|0.497855|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.998613|0.999133|0.999653|0.998267|0.999133|0.999133|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|0.628511|0.470321|0.375748|0.307464|0.518110|0.518088|
</details>
@@ -430,8 +340,8 @@ This mode contains **12 result rows**, **8 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|873411|97544|0|11263658798|97544 / 11263756342|0 / 873411|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|288864|96545|584547|11263659797|96545 / 11263756342|584547 / 873411|
|1|Radixor|ALL_CANDIDATES|858417|1490|0|10594962044|1490 / 10594963534|0 / 858417|
|2|HUNSPELL GERMAN LUCENE FILTER|ALL_CANDIDATES|284421|66639|573996|10594896895|66639 / 10594963534|573996 / 858417|
</details>
@@ -441,20 +351,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|59114|47898|49646|14978|9.978814%|8|167157|
|HUNSPELL GERMAN LUCENE FILTER|10771|23683|11866|4989|3.323828%|3|155207|
|Radixor|56726|80|1410|10454|7.181227%|8|157137|
|HUNSPELL GERMAN LUCENE FILTER|10454|18296|7735|4538|3.117315%|3|150205|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -463,16 +373,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `DE_DE`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

301
docs/benchmarks/languages/hebrew.md Normal file
View File

@@ -0,0 +1,301 @@
# Hebrew Stemmer Benchmarks
This page reports same-language stemming benchmarks for Hebrew. Accuracy is listed first because speed without root agreement is not enough to interpret stemmer quality.
All speed values are environment-specific and were measured on the hardware and JVM listed in the [benchmark overview](../index.md). Speed benchmark operations process changed dictionary tokens only. Accuracy uses the complete Radixor dictionary for the language.
The default Hebrew model currently has no same-language third-party adapter in the benchmark matrix. Its Radixor measurements are still published so the complete default-model language universe has identical corpus, command-distribution, exact-root, runtime, and pairwise-quality coverage.
## Dictionary Corpus
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `he-il-default` | `1.0.0` | `HE_IL` | 2,358 | 61,071 | 4,715 | 56,356 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **61,071**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `DeletePrefixCommand` | Deletes one or more leading characters from the word form in forward traversal. | 2,188 | 3.583% |
| `ForwardCompoundCommand` | Applies a multi-step forward patch made from skip, delete, insert, and replace operations. | 51,692 | 84.642% |
| `PrependCharacterCommand` | Prepends one character to the beginning of the word form. | 11 | 0.018% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 4,828 | 7.906% |
| `ReplaceFirstCharacterCommand` | Replaces the first character of the word form in forward traversal. | 2,352 | 3.851% |
## Accuracy
Accuracy is computed from JMH auxiliary counters in the current report. The counters are deterministic for a fixed corpus and stemmer; percentages divide matching counters by evaluated counters from the same report and are not timing metrics.
| Stemmer | All exact | Changed exact | Root preserved | Note |
| --- | ---: | ---: | ---: | --- |
| Radixor | 98.228% | 98.172% | 98.897% | Full default-model Radixor dictionary patch-command stemmer. |
## Speed
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `hebrewRadixor` | 3.921 | 0.140 | 69.6 | 1.000 | Full default-model Radixor dictionary patch-command stemmer. |
## Interpretation Notes
- Radixor is a dictionary-derived patch-command stemmer. Its quality depends on the default language model used to train the compiled trie.
- Hebrew patch commands use forward traversal as declared by the model metadata.
- Results are environment-specific and should be compared only with rows from the same benchmark run.
<!-- STEMMING-QUALITY:START -->
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `HE_IL` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The default model is `he-il-default`, loaded from classpath resource `org/egothor/stemmer/models/he-il-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.986075** among 1 deterministic stemmers; no same-language competitor was available. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.986075** among 1 deterministic stemmers; no same-language competitor was available. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **3 result rows**, **1 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.986075|0.000000%|2.784905%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.972151|1.000000|0.986075|0.999988|0.000012|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.994303|0.985879|0.977596|0.972151|0.985977|0.985971|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|685765|0|19645|1661488243|0 / 1661488243|19645 / 705410|
</details>
#### `ANY_CANDIDATE` oracle bounds
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 1661488243|0 / 705410|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|705410|0|0|1661488243|0 / 1661488243|0 / 705410|
</details>
#### Multi-output analysis
Alternative candidates are capability analyses, not replacements for the deterministic comparison.
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|19645|0|0|984|1.706615%|40|58714|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **3 result rows**, **1 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.986075|0.000000%|2.784905%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.972151|1.000000|0.986075|0.999988|0.000012|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.994303|0.985879|0.977596|0.972151|0.985977|0.985971|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|685765|0|19645|1661488243|0 / 1661488243|19645 / 705410|
</details>
#### `ANY_CANDIDATE` oracle bounds
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 1661488243|0 / 705410|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|705410|0|0|1661488243|0 / 1661488243|0 / 705410|
</details>
#### Multi-output analysis
Alternative candidates are capability analyses, not replacements for the deterministic comparison.
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|19645|0|0|984|1.706615%|40|58714|
### Output Policies and Metric Definitions
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
- Jaccard index: `TP / (TP + FP + FN)`.
- Fowlkes–Mallows index: `sqrt(precision * recall)`.
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `HE_IL`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

293
docs/benchmarks/languages/hungarian.md
View File

@@ -8,20 +8,20 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `HU_HU` | 19,406 | 935,713 | 38,775 | 896,938 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `hu-hu-default` | `1.0.0` | `HU_HU` | 19,406 | 935,713 | 38,775 | 896,938 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **935,713**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **935,713**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 15 | 0.002% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 149,173 | 15.942% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 746,296 | 79.757% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 40,125 | 4.288% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 750,282 | 80.183% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 36,139 | 3.862% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 104 | 0.011% |
## Accuracy
@@ -35,16 +35,22 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 66.445% | 66.938% | 55.043% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene HungarianLightStemFilter | 14.748% | 14.777% | 14.086% | Light suffix stemmer; intentionally narrower than a dictionary-derived stemmer. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `hungarianRadixor` | 62.232 | 6.412 | 69.4 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HungarianLightStemFilter | `hungarianLuceneHungarianLightStemFilter` | 92.813 | 6.929 | 103.5 | 1.491 | Light Hungarian suffix stemmer. |
| Official Snowball direct | `snowballDirect[HUNGARIAN]` | 157.765 | 13.202 | 175.9 | 2.535 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[HUNGARIAN]` | 188.863 | 15.880 | 210.6 | 3.035 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `hungarianRadixor` | 61.205 | 0.944 | 68.2 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HungarianLightStemFilter | `hungarianLuceneHungarianLightStemFilter` | 92.090 | 3.410 | 102.7 | 1.505 | Light Hungarian suffix stemmer. |
| Official Snowball direct | `snowballDirect[HUNGARIAN]` | 152.969 | 4.468 | 170.5 | 2.499 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[HUNGARIAN]` | 188.807 | 5.290 | 210.5 | 3.085 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -58,30 +64,30 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `HU_HU` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `HU_HU` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/hu_hu/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `hu-hu-default`, loaded from classpath resource `org/egothor/stemmer/models/hu-hu-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.995491** among 4 deterministic stemmers. The runner-up is `SNOWBALL HUNGARIAN LUCENE FILTER` at 0.822606, a difference of 0.172885. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.996163** among 4 deterministic stemmers. The runner-up is `SNOWBALL HUNGARIAN DIRECT` at 0.821708, a difference of 0.174455. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.995555** among 4 deterministic stemmers. The runner-up is `SNOWBALL HUNGARIAN LUCENE FILTER` at 0.822963, a difference of 0.172592. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.996227** among 4 deterministic stemmers. The runner-up is `SNOWBALL HUNGARIAN DIRECT` at 0.822077, a difference of 0.174151. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.995491|272900 / 419820542893 (0.000065%)|199837 / 22162103 (0.901706%)|0.988376|0.989352|0.989353|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.822606|1792049 / 419820542893 (0.000427%)|7862745 / 22162103 (35.478334%)|0.826288|0.747610|0.757196|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.822348|1506056 / 419820542893 (0.000359%)|7874191 / 22162103 (35.529981%)|0.837137|0.752866|0.763681|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.816668|4132555 / 419820542893 (0.000984%)|8125833 / 22162103 (36.665442%)|0.740018|0.696055|0.699478|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.995555|&lt;0.000001%|0.889037%|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|0.822963|0.000378%|35.407050%|
|3|SNOWBALL HUNGARIAN DIRECT|0.822704|0.000309%|35.458800%|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|0.816967|0.000915%|36.605593%|
</div>
@@ -89,10 +95,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987727|0.990983|0.999999|0.995491|0.999999|0.000001|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.888633|0.645217|0.999996|0.822606|0.999977|0.000023|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.904644|0.644700|0.999996|0.822348|0.999978|0.000022|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.772547|0.633346|0.999990|0.816668|0.999971|0.000029|
|1|Radixor|PRIMARY_OUTPUT|0.999998|0.991110|1.000000|0.995555|1.000000|0.000000|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.901236|0.645929|0.999996|0.822963|0.999977|0.000023|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.917622|0.645412|0.999997|0.822704|0.999978|0.000022|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.786953|0.633944|0.999991|0.816967|0.999971|0.000029|
</details>
@@ -100,21 +106,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988376|0.989352|0.990330|0.978929|0.989353|0.989353|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.826288|0.747610|0.682613|0.596947|0.757206|0.757196|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.837137|0.752866|0.684009|0.603677|0.763691|0.763681|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.740018|0.696055|0.657023|0.533807|0.699492|0.699478|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989352|0.998036|0.997809|0.997922|0.997922|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.747599|0.990687|0.924490|0.956445|0.956445|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.752855|0.991948|0.924304|0.956932|0.956932|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.696040|0.982615|0.926772|0.953877|0.953877|
|1|Radixor|PRIMARY_OUTPUT|0.998208|0.995534|0.992875|0.991108|0.995544|0.995544|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.835212|0.752518|0.684724|0.603229|0.762978|0.762967|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.846240|0.757814|0.686119|0.610064|0.769574|0.769564|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.750715|0.702210|0.659593|0.541082|0.706318|0.706304|
</details>
@@ -122,62 +117,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|21962266|272900|199837|419820269993|272900 / 419820542893|199837 / 22162103|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|14299358|1792049|7862745|419818750844|1792049 / 419820542893|7862745 / 22162103|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|14287912|1506056|7874191|419819036837|1506056 / 419820542893|7874191 / 22162103|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|14036270|4132555|8125833|419816410338|4132555 / 419820542893|8125833 / 22162103|
|1|Radixor|PRIMARY_OUTPUT|21921219|39|196636|414653743434|39 / 414653743473|196636 / 22117855|
|2|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|14286575|1565633|7831280|414652177840|1565633 / 414653743473|7831280 / 22117855|
|3|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|14275129|1281527|7842726|414652461946|1281527 / 414653743473|7842726 / 22117855|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|14021483|3795942|8096372|414649947531|3795942 / 414653743473|8096372 / 22117855|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 419820542893 (0.000000%)|0 / 22162103 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|22162103|0|0|419820542893|0 / 419820542893|0 / 22162103|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 414653743473|0 / 22117855|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|460158 / 419820542893 (0.000110%)|0 / 22162103 (0.000000%)|0.983661|0.989725|0.989777|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
</div>
@@ -185,7 +158,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.979659|1.000000|0.999999|0.999999|0.999999|0.000001|
|1|Radixor|ALL_CANDIDATES|0.999991|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -193,15 +166,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.983661|0.989725|0.995865|0.979659|0.989777|0.989777|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999993|0.999996|0.999998|0.999991|0.999996|0.999996|
</details>
@@ -209,7 +174,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|22162103|460158|0|419820082735|460158 / 419820542893|0 / 22162103|
|1|Radixor|ALL_CANDIDATES|22117855|192|0|414653743281|192 / 414653743473|0 / 22117855|
</details>
@@ -219,22 +184,22 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|199837|272900|187258|12320|1.344473%|5|929326|
|Radixor|196636|39|153|6664|0.731754%|5|917595|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996163|272775 / 385870694917 (0.000071%)|164277 / 21411411 (0.767240%)|0.988321|0.989820|0.989822|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.821708|1496670 / 385870694917 (0.000388%)|7634885 / 21411411 (35.658019%)|0.834899|0.751079|0.761809|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.821708|1496670 / 385870694917 (0.000388%)|7634885 / 21411411 (35.658019%)|0.834899|0.751079|0.761809|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.815077|3639046 / 385870694917 (0.000943%)|7918708 / 21411411 (36.983588%)|0.750108|0.700135|0.704477|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.996227|&lt;0.000001%|0.754564%|
|2|SNOWBALL HUNGARIAN DIRECT|0.822077|0.000334%|35.584346%|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|0.822077|0.000334%|35.584346%|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|0.815385|0.000869%|36.922109%|
</div>
@@ -242,10 +207,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987325|0.992328|0.999999|0.996163|0.999999|0.000001|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.902007|0.643420|0.999996|0.821708|0.999976|0.000024|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.902007|0.643420|0.999996|0.821708|0.999976|0.000024|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.787585|0.630164|0.999991|0.815077|0.999970|0.000030|
|1|Radixor|PRIMARY_OUTPUT|0.999998|0.992454|1.000000|0.996227|1.000000|0.000000|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.915319|0.644157|0.999997|0.822077|0.999977|0.000023|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.915319|0.644157|0.999997|0.822077|0.999977|0.000023|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.802756|0.630779|0.999991|0.815385|0.999971|0.000029|
</details>
@@ -253,21 +218,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988321|0.989820|0.991323|0.979845|0.989823|0.989822|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.834899|0.751079|0.682555|0.601383|0.761820|0.761809|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.834899|0.751079|0.682555|0.601383|0.761820|0.761809|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.750108|0.700135|0.656404|0.538621|0.704491|0.704477|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989819|0.997945|0.998273|0.998109|0.998109|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.751068|0.991610|0.923288|0.956230|0.956230|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.751068|0.991610|0.923288|0.956230|0.956230|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.700120|0.983687|0.925487|0.953700|0.953700|
|1|Radixor|PRIMARY_OUTPUT|0.998480|0.996212|0.993954|0.992453|0.996219|0.996219|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|0.844241|0.756163|0.684726|0.607928|0.767860|0.767849|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|0.844241|0.756163|0.684726|0.607928|0.767860|0.767849|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.761246|0.706452|0.659016|0.546135|0.711591|0.711577|
</details>
@@ -275,62 +229,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|21247134|272775|164277|385870422142|272775 / 385870694917|164277 / 21411411|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|13776526|1496670|7634885|385869198247|1496670 / 385870694917|7634885 / 21411411|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|13776526|1496670|7634885|385869198247|1496670 / 385870694917|7634885 / 21411411|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|13492703|3639046|7918708|385867055871|3639046 / 385870694917|7918708 / 21411411|
|1|Radixor|PRIMARY_OUTPUT|21206087|39|161230|380936197647|39 / 380936197686|161230 / 21367317|
|2|SNOWBALL HUNGARIAN DIRECT|PRIMARY_OUTPUT|13763897|1273370|7603420|380934924316|1273370 / 380936197686|7603420 / 21367317|
|3|SNOWBALL HUNGARIAN LUCENE FILTER|PRIMARY_OUTPUT|13763897|1273370|7603420|380934924316|1273370 / 380936197686|7603420 / 21367317|
|4|HUNGARIAN LUCENE HUNGARIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|13478053|3311675|7889264|380932886011|3311675 / 380936197686|7889264 / 21367317|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 385870694917 (0.000000%)|0 / 21411411 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|21411411|0|0|385870694917|0 / 385870694917|0 / 21411411|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 380936197686|0 / 21367317|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|458462 / 385870694917 (0.000119%)|0 / 21411411 (0.000000%)|0.983159|0.989407|0.989462|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
</div>
@@ -338,7 +270,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.979037|1.000000|0.999999|0.999999|0.999999|0.000001|
|1|Radixor|ALL_CANDIDATES|0.999991|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -346,15 +278,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.983159|0.989407|0.995736|0.979037|0.989463|0.989462|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999993|0.999996|0.999998|0.999991|0.999996|0.999996|
</details>
@@ -362,7 +286,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|21411411|458462|0|385870236455|458462 / 385870694917|0 / 21411411|
|1|Radixor|ALL_CANDIDATES|21367317|192|0|380936197494|192 / 380936197686|0 / 21367317|
</details>
@@ -372,19 +296,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|164277|272775|185687|11153|1.269532%|5|890245|
|Radixor|161230|39|153|5518|0.632162%|5|878574|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -393,16 +317,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `HU_HU`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

3
docs/benchmarks/languages/index.md
View File

@@ -1,6 +1,6 @@
# Language Benchmark Pages
This section splits Radixor stemmer benchmark results by language. Each language page preserves the existing exact-root accuracy and runtime-performance results and adds pairwise stemming-quality tables for both dictionary-processing modes.
This section splits Radixor stemmer benchmark results by language. Each of the 20 registered default models has one language page containing the refreshed corpus, patch-command distribution, exact-root accuracy, runtime performance, and pairwise stemming-quality tables for both dictionary-processing modes.
## Reference Pages
@@ -23,6 +23,7 @@ This section splits Radixor stemmer benchmark results by language. Each language
| Finnish | `FI_FI` | [Finnish](finnish.md) |
| French | `FR_FR` | [French](french.md) |
| German | `DE_DE` | [German](german.md) |
| Hebrew | `HE_IL` | [Hebrew](hebrew.md) |
| Hungarian | `HU_HU` | [Hungarian](hungarian.md) |
| Italian | `IT_IT` | [Italian](italian.md) |
| Norwegian Bokmal | `NB_NO` | [Norwegian Bokmal](norwegian-bokmal.md) |

293
docs/benchmarks/languages/italian.md
View File

@@ -8,20 +8,20 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `IT_IT` | 10,009 | 337,546 | 20,004 | 317,542 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `it-it-default` | `1.0.0` | `IT_IT` | 10,009 | 337,546 | 20,004 | 317,542 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **337,546**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **337,546**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 302,171 | 89.520% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 302,089 | 89.496% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 12,348 | 3.658% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 20,013 | 5.929% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 3,014 | 0.893% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 3,096 | 0.917% |
## Accuracy
@@ -34,16 +34,22 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene SnowballFilter | 0.041% | 0.043% | 0.010% | Lucene TokenFilter integration path around the Snowball algorithm. |
| Official Snowball direct | 0.041% | 0.043% | 0.010% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `italianRadixor` | 24.491 | 3.128 | 77.1 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene ItalianLightStemFilter | `italianLuceneItalianLightStemFilter` | 15.977 | 1.041 | 50.3 | 0.652 | Light Italian suffix stemmer. |
| Official Snowball direct | `snowballDirect[ITALIAN]` | 109.526 | 12.572 | 344.9 | 4.472 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[ITALIAN]` | 116.260 | 7.459 | 366.1 | 4.747 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `italianRadixor` | 25.073 | 0.534 | 79.0 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene ItalianLightStemFilter | `italianLuceneItalianLightStemFilter` | 15.956 | 0.184 | 50.2 | 0.636 | Light Italian suffix stemmer. |
| Official Snowball direct | `snowballDirect[ITALIAN]` | 115.818 | 3.174 | 364.7 | 4.619 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[ITALIAN]` | 123.974 | 4.405 | 390.4 | 4.944 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -57,30 +63,30 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `IT_IT` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `IT_IT` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/it_it/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `it-it-default`, loaded from classpath resource `org/egothor/stemmer/models/it-it-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.996507** among 4 deterministic stemmers. The runner-up is `SNOWBALL ITALIAN DIRECT` at 0.866189, a difference of 0.130318. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.996512** among 4 deterministic stemmers. The runner-up is `SNOWBALL ITALIAN DIRECT` at 0.866205, a difference of 0.130307. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.996651** among 4 deterministic stemmers. The runner-up is `SNOWBALL ITALIAN DIRECT` at 0.866290, a difference of 0.130361. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.996656** among 4 deterministic stemmers. The runner-up is `SNOWBALL ITALIAN DIRECT` at 0.866307, a difference of 0.130350. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996507|124172 / 53638521211 (0.000231%)|42908 / 6143814 (0.698394%)|0.982618|0.986492|0.986512|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.866189|504775 / 53638521211 (0.000941%)|1644164 / 6143814 (26.761292%)|0.859975|0.807240|0.811470|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.866189|504775 / 53638521211 (0.000941%)|1644164 / 6143814 (26.761292%)|0.859975|0.807240|0.811470|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.508926|10589 / 53638521211 (0.000020%)|6034130 / 6143814 (98.214725%)|0.082782|0.035020|0.127588|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.996651|0.000000%|0.669827%|
|2|SNOWBALL ITALIAN DIRECT|0.866290|0.000738%|26.741219%|
|3|SNOWBALL ITALIAN LUCENE FILTER|0.866290|0.000738%|26.741219%|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|0.508920|0.000005%|98.216094%|
</div>
@@ -88,10 +94,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.980053|0.993016|0.999998|0.996507|0.999997|0.000003|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.899134|0.732387|0.999991|0.866189|0.999960|0.000040|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.899134|0.732387|0.999991|0.866189|0.999960|0.000040|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.911959|0.017853|1.000000|0.508926|0.999887|0.000113|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.993302|1.000000|0.996651|0.999999|0.000001|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.920474|0.732588|0.999993|0.866290|0.999961|0.000039|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.920474|0.732588|0.999993|0.866290|0.999961|0.000039|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.975468|0.017839|1.000000|0.508920|0.999885|0.000115|
</details>
@@ -99,21 +105,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.982618|0.986492|0.990396|0.973344|0.986513|0.986512|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.859975|0.807240|0.760598|0.676783|0.811489|0.811470|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.859975|0.807240|0.760598|0.676783|0.811489|0.811470|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.082782|0.035020|0.022207|0.017822|0.127597|0.127588|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986490|0.995780|0.997113|0.996446|0.996446|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.807220|0.987994|0.933408|0.959925|0.959925|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.807220|0.987994|0.933408|0.959925|0.959925|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.035016|0.997481|0.737537|0.848037|0.848037|
|1|Radixor|PRIMARY_OUTPUT|0.998653|0.996640|0.994634|0.993302|0.996645|0.996645|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.875563|0.815854|0.763768|0.688980|0.821175|0.821157|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.875563|0.815854|0.763768|0.688980|0.821175|0.821157|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.083115|0.035037|0.022197|0.017831|0.131914|0.131907|
</details>
@@ -121,62 +116,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|6100906|124172|42908|53638397039|124172 / 53638521211|42908 / 6143814|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|4499650|504775|1644164|53638016436|504775 / 53638521211|1644164 / 6143814|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|4499650|504775|1644164|53638016436|504775 / 53638521211|1644164 / 6143814|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|109684|10589|6034130|53638510622|10589 / 53638521211|6034130 / 6143814|
|1|Radixor|PRIMARY_OUTPUT|6093034|0|41088|52600354673|0 / 52600354673|41088 / 6134122|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|4493783|388246|1640339|52599966427|388246 / 52600354673|1640339 / 6134122|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|4493783|388246|1640339|52599966427|388246 / 52600354673|1640339 / 6134122|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|109427|2752|6024695|52600351921|2752 / 52600354673|6024695 / 6134122|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999993|0 / 53638521211 (0.000000%)|80 / 6143814 (0.001302%)|0.999997|0.999993|0.999993|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.001304%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0.999987|1.000000|0.999993|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999997|0.999993|0.999990|0.999987|0.999993|0.999993|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|6143734|0|80|53638521211|0 / 53638521211|80 / 6143814|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 52600354673|80 / 6134122|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999992|170950 / 53638521211 (0.000319%)|80 / 6143814 (0.001302%)|0.978222|0.986272|0.986363|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.999993|0.000000%|0.001304%|
</div>
@@ -184,7 +157,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.972928|0.999987|0.999997|0.999992|0.999997|0.000003|
|1|Radixor|ALL_CANDIDATES|1.000000|0.999987|1.000000|0.999993|1.000000|0.000000|
</details>
@@ -192,15 +165,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.978222|0.986272|0.994455|0.972916|0.986365|0.986363|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999997|0.999993|0.999990|0.999987|0.999993|0.999993|
</details>
@@ -208,7 +173,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|6143734|170950|80|53638350261|170950 / 53638521211|80 / 6143814|
|1|Radixor|ALL_CANDIDATES|6134042|0|80|52600354673|0 / 52600354673|80 / 6134122|
</details>
@@ -218,22 +183,22 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|42828|124172|46778|6254|1.909321%|4|334175|
|Radixor|41008|0|0|3069|0.946153%|4|327552|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996512|124171 / 53611667072 (0.000232%)|42828 / 6142174 (0.697278%)|0.982617|0.986495|0.986515|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.866205|504774 / 53611667072 (0.000942%)|1643522 / 6142174 (26.757985%)|0.859970|0.807252|0.811479|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.866205|504774 / 53611667072 (0.000942%)|1643522 / 6142174 (26.757985%)|0.859970|0.807252|0.811479|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.508927|10588 / 53611667072 (0.000020%)|6032516 / 6142174 (98.214671%)|0.082784|0.035021|0.127589|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.996656|0.000000%|0.668702%|
|2|SNOWBALL ITALIAN DIRECT|0.866307|0.000738%|26.737902%|
|3|SNOWBALL ITALIAN LUCENE FILTER|0.866307|0.000738%|26.737902%|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|0.508920|0.000005%|98.216040%|
</div>
@@ -241,10 +206,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.980048|0.993027|0.999998|0.996512|0.999997|0.000003|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.899114|0.732420|0.999991|0.866205|0.999960|0.000040|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.899114|0.732420|0.999991|0.866205|0.999960|0.000040|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.911947|0.017853|1.000000|0.508927|0.999887|0.000113|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.993313|1.000000|0.996656|0.999999|0.000001|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.920458|0.732621|0.999993|0.866307|0.999961|0.000039|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.920458|0.732621|0.999993|0.866307|0.999961|0.000039|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.975462|0.017840|1.000000|0.508920|0.999885|0.000115|
</details>
@@ -252,21 +217,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.982617|0.986495|0.990404|0.973350|0.986516|0.986515|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.859970|0.807252|0.760624|0.676800|0.811498|0.811479|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.859970|0.807252|0.760624|0.676800|0.811498|0.811479|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.082784|0.035021|0.022208|0.017823|0.127598|0.127589|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986493|0.995780|0.997115|0.996447|0.996447|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.807232|0.987991|0.933413|0.959927|0.959927|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.807232|0.987991|0.933413|0.959927|0.959927|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.035017|0.997481|0.737534|0.848035|0.848035|
|1|Radixor|PRIMARY_OUTPUT|0.998655|0.996645|0.994643|0.993313|0.996651|0.996650|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|0.875561|0.815868|0.763794|0.689001|0.821186|0.821168|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|0.875561|0.815868|0.763794|0.689001|0.821186|0.821168|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.083118|0.035038|0.022198|0.017832|0.131916|0.131908|
</details>
@@ -274,62 +228,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|6099346|124171|42828|53611542901|124171 / 53611667072|42828 / 6142174|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|4498652|504774|1643522|53611162298|504774 / 53611667072|1643522 / 6142174|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|4498652|504774|1643522|53611162298|504774 / 53611667072|1643522 / 6142174|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|109658|10588|6032516|53611656484|10588 / 53611667072|6032516 / 6142174|
|1|Radixor|PRIMARY_OUTPUT|6091474|0|41008|52574085988|0 / 52574085988|41008 / 6132482|
|2|SNOWBALL ITALIAN DIRECT|PRIMARY_OUTPUT|4492785|388246|1639697|52573697742|388246 / 52574085988|1639697 / 6132482|
|3|SNOWBALL ITALIAN LUCENE FILTER|PRIMARY_OUTPUT|4492785|388246|1639697|52573697742|388246 / 52574085988|1639697 / 6132482|
|4|ITALIAN LUCENE ITALIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|109401|2752|6023081|52574083236|2752 / 52574085988|6023081 / 6132482|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 53611667072 (0.000000%)|0 / 6142174 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|6142174|0|0|53611667072|0 / 53611667072|0 / 6142174|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 52574085988|0 / 6132482|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999998|170949 / 53611667072 (0.000319%)|0 / 6142174 (0.000000%)|0.978219|0.986275|0.986366|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -337,7 +269,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.972922|1.000000|0.999997|0.999998|0.999997|0.000003|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -345,15 +277,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.978219|0.986275|0.994464|0.972922|0.986368|0.986366|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -361,7 +285,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|6142174|170949|0|53611496123|170949 / 53611667072|0 / 6142174|
|1|Radixor|ALL_CANDIDATES|6132482|0|0|52574085988|0 / 52574085988|0 / 6132482|
</details>
@@ -371,19 +295,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|42828|124171|46778|6252|1.909188%|4|334089|
|Radixor|41008|0|0|3068|0.946081%|4|327469|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -392,16 +316,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `IT_IT`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

319
docs/benchmarks/languages/norwegian-bokmal.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `NB_NO` | 17,929 | 90,757 | 33,376 | 57,381 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `nb-no-default` | `1.0.0` | `NB_NO` | 17,929 | 90,757 | 33,376 | 57,381 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **90,757**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **90,757**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 1,500 | 1.653% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 4,296 | 4.734% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 47,619 | 52.469% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 34,420 | 37.925% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 2,922 | 3.220% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 2,528 | 2.785% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 4,258 | 4.692% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 48,925 | 53.908% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 32,086 | 35.354% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 2,960 | 3.261% |
## Accuracy
@@ -36,17 +36,23 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene SnowballFilter | 54.803% | 51.780% | 60.001% | Lucene TokenFilter integration path around the Snowball algorithm. |
| Lucene NorwegianLightStemFilter | 52.136% | 50.616% | 54.749% | Light suffix stemmer; intentionally narrower than a dictionary-derived stemmer. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `norwegianBokmalRadixor` | 3.631 | 1.377 | 63.3 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene NorwegianMinimalStemFilter | `norwegianBokmalLuceneNorwegianMinimalStemFilter` | 2.910 | 0.177 | 50.7 | 0.801 | Minimal Norwegian suffix reducer. |
| Lucene NorwegianLightStemFilter | `norwegianBokmalLuceneNorwegianLightStemFilter` | 3.335 | 0.116 | 58.1 | 0.919 | Light Norwegian suffix stemmer. |
| Official Snowball direct | `snowballDirect[NORWEGIAN_BOKMAL]` | 4.277 | 0.082 | 74.5 | 1.178 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[NORWEGIAN_BOKMAL]` | 6.077 | 0.208 | 105.9 | 1.674 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `norwegianBokmalRadixor` | 3.401 | 0.055 | 59.3 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene NorwegianMinimalStemFilter | `norwegianBokmalLuceneNorwegianMinimalStemFilter` | 2.943 | 0.023 | 51.3 | 0.865 | Minimal Norwegian suffix reducer. |
| Lucene NorwegianLightStemFilter | `norwegianBokmalLuceneNorwegianLightStemFilter` | 3.358 | 0.036 | 58.5 | 0.987 | Light Norwegian suffix stemmer. |
| Official Snowball direct | `snowballDirect[NORWEGIAN_BOKMAL]` | 4.378 | 0.295 | 76.3 | 1.287 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[NORWEGIAN_BOKMAL]` | 6.114 | 0.436 | 106.5 | 1.797 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -60,31 +66,31 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `NB_NO` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `NB_NO` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/nb_no/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `nb-no-default`, loaded from classpath resource `org/egothor/stemmer/models/nb-no-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.974783** among 5 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN BOKMAL DIRECT` at 0.874964, a difference of 0.099819. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.975000** among 5 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN BOKMAL DIRECT` at 0.874991, a difference of 0.100009. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.976021** among 5 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN BOKMAL DIRECT` at 0.874259, a difference of 0.101762. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.976240** among 5 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN BOKMAL DIRECT` at 0.874286, a difference of 0.101954. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.974783|11482 / 2835618215 (0.000405%)|7170 / 142180 (5.042903%)|0.927078|0.935387|0.935488|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.874964|23997 / 2835618215 (0.000846%)|35554 / 142180 (25.006330%)|0.802095|0.781707|0.782399|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.874834|24046 / 2835618215 (0.000848%)|35591 / 142180 (25.032353%)|0.801759|0.781401|0.782091|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.850006|25171 / 2835618215 (0.000888%)|42651 / 142180 (29.997890%)|0.776381|0.745871|0.747464|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.832414|14772 / 2835618215 (0.000521%)|47654 / 142180 (33.516669%)|0.815763|0.751764|0.758263|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.976021|0.000000%|4.795770%|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|0.874259|0.000386%|25.147805%|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|0.874138|0.000389%|25.171937%|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|0.849389|0.000416%|30.121722%|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|0.831282|0.000110%|33.743568%|
</div>
@@ -92,11 +98,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.921620|0.949571|0.999996|0.974783|0.999993|0.000007|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.816288|0.749937|0.999992|0.874964|0.999979|0.000021|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.815930|0.749676|0.999992|0.874834|0.999979|0.000021|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.798148|0.700021|0.999991|0.850006|0.999976|0.000024|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.864847|0.664833|0.999995|0.832414|0.999978|0.000022|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.952042|1.000000|0.976021|0.999997|0.000003|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.910734|0.748522|0.999996|0.874259|0.999983|0.000017|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.910189|0.748281|0.999996|0.874138|0.999983|0.000017|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.898501|0.698783|0.999996|0.849389|0.999980|0.000020|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.969387|0.662564|0.999999|0.831282|0.999981|0.000019|
</details>
@@ -104,23 +110,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.927078|0.935387|0.943846|0.878617|0.935491|0.935488|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.802095|0.781707|0.762330|0.641641|0.782409|0.782399|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.801759|0.781401|0.762052|0.641229|0.782102|0.782091|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.776381|0.745871|0.717668|0.594732|0.747476|0.747464|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.815763|0.751764|0.697076|0.602261|0.758274|0.758263|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.935384|0.993354|0.994615|0.993984|0.993984|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.781696|0.988328|0.971120|0.979648|0.979648|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.781391|0.988295|0.971086|0.979615|0.979615|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.745859|0.987774|0.965622|0.976573|0.976573|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.751753|0.992089|0.962516|0.977079|0.977079|
|1|Radixor|PRIMARY_OUTPUT|0.990026|0.975432|0.961262|0.952042|0.975727|0.975725|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.872901|0.821699|0.776171|0.697359|0.825654|0.825646|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.872435|0.821332|0.775884|0.696830|0.825274|0.825266|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.849918|0.786156|0.731293|0.647658|0.792374|0.792365|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.887216|0.787133|0.707341|0.648985|0.801425|0.801417|
</details>
@@ -128,63 +122,41 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|135010|11482|7170|2835606733|11482 / 2835618215|7170 / 142180|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|106626|23997|35554|2835594218|23997 / 2835618215|35554 / 142180|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|106589|24046|35591|2835594169|24046 / 2835618215|35591 / 142180|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|99529|25171|42651|2835593044|25171 / 2835618215|42651 / 142180|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|94526|14772|47654|2835603443|14772 / 2835618215|47654 / 142180|
|1|Radixor|PRIMARY_OUTPUT|134138|0|6757|2676746970|0 / 2676746970|6757 / 140895|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|105463|10337|35432|2676736633|10337 / 2676746970|35432 / 140895|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|105429|10403|35466|2676736567|10403 / 2676746970|35466 / 140895|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|98455|11122|42440|2676735848|11122 / 2676746970|42440 / 140895|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|93352|2948|47543|2676744022|2948 / 2676746970|47543 / 140895|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 2835618215 (0.000000%)|0 / 142180 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|142180|0|0|2835618215|0 / 2835618215|0 / 142180|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 2676746970|0 / 140895|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999996|20161 / 2835618215 (0.000711%)|0 / 142180 (0.000000%)|0.898118|0.933794|0.935844|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -192,7 +164,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.875811|1.000000|0.999993|0.999996|0.999993|0.000007|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -200,15 +172,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.898118|0.933794|0.972422|0.875811|0.935848|0.935844|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -216,7 +180,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|142180|20161|0|2835598054|20161 / 2835618215|0 / 142180|
|1|Radixor|ALL_CANDIDATES|140895|0|0|2676746970|0 / 2676746970|0 / 140895|
</details>
@@ -226,23 +190,23 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|7170|11482|8679|4237|5.626079%|9|79825|
|Radixor|6757|0|0|2097|2.865929%|9|75343|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.975000|11482 / 2831176784 (0.000406%)|7104 / 142091 (4.999613%)|0.927151|0.935591|0.935695|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.874991|23997 / 2831176784 (0.000848%)|35524 / 142091 (25.000880%)|0.802043|0.781698|0.782388|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.874798|23993 / 2831176784 (0.000847%)|35579 / 142091 (25.039587%)|0.801914|0.781464|0.782161|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.849947|25118 / 2831176784 (0.000887%)|42641 / 142091 (30.009642%)|0.776513|0.745896|0.747500|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.832344|14719 / 2831176784 (0.000520%)|47644 / 142091 (33.530625%)|0.815950|0.751796|0.758325|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.976240|0.000000%|4.751928%|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|0.874286|0.000387%|25.142395%|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|0.874101|0.000387%|25.179325%|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|0.849330|0.000414%|30.133659%|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|0.831210|0.000108%|33.757794%|
</div>
@@ -250,11 +214,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.921608|0.950004|0.999996|0.975000|0.999993|0.000007|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.816205|0.749991|0.999992|0.874991|0.999979|0.000021|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.816153|0.749604|0.999992|0.874798|0.999979|0.000021|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.798359|0.699904|0.999991|0.849947|0.999976|0.000024|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.865169|0.664694|0.999995|0.832344|0.999978|0.000022|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.952481|1.000000|0.976240|0.999997|0.000003|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.910689|0.748576|0.999996|0.874286|0.999983|0.000017|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.910546|0.748207|0.999996|0.874101|0.999983|0.000017|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.898862|0.698663|0.999996|0.849330|0.999980|0.000020|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.969896|0.662422|0.999999|0.831210|0.999981|0.000019|
</details>
@@ -262,23 +226,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.927151|0.935591|0.944186|0.878976|0.935698|0.935695|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.802043|0.781698|0.762360|0.641630|0.782398|0.782388|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.801914|0.781464|0.762031|0.641314|0.782171|0.782161|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.776513|0.745896|0.717603|0.594765|0.747512|0.747500|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.815950|0.751796|0.696995|0.602302|0.758335|0.758325|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.935587|0.993348|0.994694|0.994020|0.994020|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.781688|0.988318|0.971127|0.979647|0.979647|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.781454|0.988310|0.971074|0.979616|0.979616|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.745885|0.987789|0.965603|0.976570|0.976570|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.751785|0.992107|0.962494|0.977076|0.977076|
|1|Radixor|PRIMARY_OUTPUT|0.990121|0.975662|0.961620|0.952481|0.975951|0.975950|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|0.872882|0.821713|0.776211|0.697379|0.825663|0.825655|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|0.872677|0.821432|0.775872|0.696975|0.825395|0.825387|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.850142|0.786219|0.731236|0.647743|0.792466|0.792457|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.887506|0.787200|0.707265|0.649077|0.801549|0.801541|
</details>
@@ -286,63 +238,41 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|134987|11482|7104|2831165302|11482 / 2831176784|7104 / 142091|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|106567|23997|35524|2831152787|23997 / 2831176784|35524 / 142091|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|106512|23993|35579|2831152791|23993 / 2831176784|35579 / 142091|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|99450|25118|42641|2831151666|25118 / 2831176784|42641 / 142091|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|94447|14719|47644|2831162065|14719 / 2831176784|47644 / 142091|
|1|Radixor|PRIMARY_OUTPUT|134115|0|6691|2672431799|0 / 2672431799|6691 / 140806|
|2|SNOWBALL NORWEGIAN BOKMAL DIRECT|PRIMARY_OUTPUT|105404|10337|35402|2672421462|10337 / 2672431799|35402 / 140806|
|3|SNOWBALL NORWEGIAN BOKMAL LUCENE FILTER|PRIMARY_OUTPUT|105352|10350|35454|2672421449|10350 / 2672431799|35454 / 140806|
|4|NORWEGIAN BOKMAL LUCENE NORWEGIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|98376|11069|42430|2672420730|11069 / 2672431799|42430 / 140806|
|5|NORWEGIAN BOKMAL LUCENE NORWEGIAN MINIMAL STEM FILTER|PRIMARY_OUTPUT|93273|2895|47533|2672428904|2895 / 2672431799|47533 / 140806|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 2831176784 (0.000000%)|0 / 142091 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|142091|0|0|2831176784|0 / 2831176784|0 / 142091|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 2672431799|0 / 140806|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999996|20161 / 2831176784 (0.000712%)|0 / 142091 (0.000000%)|0.898061|0.933756|0.935808|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -350,7 +280,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.875743|1.000000|0.999993|0.999996|0.999993|0.000007|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -358,15 +288,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.898061|0.933756|0.972405|0.875743|0.935811|0.935808|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -374,7 +296,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|142091|20161|0|2831156623|20161 / 2831176784|0 / 142091|
|1|Radixor|ALL_CANDIDATES|140806|0|0|2672431799|0 / 2672431799|0 / 140806|
</details>
@@ -384,19 +306,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|7104|11482|8679|4204|5.586637%|9|79733|
|Radixor|6691|0|0|2064|2.823105%|9|75251|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -405,16 +327,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `NB_NO`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

279
docs/benchmarks/languages/norwegian-nynorsk.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `NN_NO` | 4,688 | 19,651 | 6,089 | 13,562 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `nn-no-default` | `1.0.0` | `NN_NO` | 4,688 | 19,651 | 6,089 | 13,562 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **19,651**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **19,651**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 224 | 1.140% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 1,505 | 7.659% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 11,017 | 56.063% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 6,427 | 32.706% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 478 | 2.432% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 312 | 1.588% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 1,456 | 7.409% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 11,325 | 57.631% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 6,031 | 30.691% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 527 | 2.682% |
## Accuracy
@@ -34,15 +34,21 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 60.974% | 60.212% | 62.670% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene SnowballFilter | 60.918% | 60.146% | 62.638% | Lucene TokenFilter integration path around the Snowball algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `radixor[NORWEGIAN_NYNORSK]` | 0.571 | 0.012 | 42.1 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Official Snowball direct | `snowballDirect[NORWEGIAN_NYNORSK]` | 0.919 | 0.038 | 67.7 | 1.609 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[NORWEGIAN_NYNORSK]` | 1.309 | 0.015 | 96.5 | 2.292 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `radixor[NORWEGIAN_NYNORSK]` | 0.617 | 0.062 | 45.5 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Official Snowball direct | `snowballDirect[NORWEGIAN_NYNORSK]` | 0.955 | 0.076 | 70.4 | 1.548 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[NORWEGIAN_NYNORSK]` | 1.352 | 0.106 | 99.7 | 2.191 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -56,29 +62,29 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `NN_NO` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `NN_NO` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/nn_no/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `nn-no-default`, loaded from classpath resource `org/egothor/stemmer/models/nn-no-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.935777** among 3 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN NYNORSK DIRECT` at 0.858908, a difference of 0.076869. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.935853** among 3 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN NYNORSK DIRECT` at 0.859037, a difference of 0.076816. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.950991** among 3 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN NYNORSK DIRECT` at 0.868094, a difference of 0.082897. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.951104** among 3 deterministic stemmers. The runner-up is `SNOWBALL NORWEGIAN NYNORSK DIRECT` at 0.868252, a difference of 0.082852. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.935777|6230 / 166491473 (0.003742%)|3936 / 30652 (12.840924%)|0.822355|0.840152|0.840669|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.858908|8274 / 166491473 (0.004970%)|8648 / 30652 (28.213493%)|0.724941|0.722271|0.722234|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.858484|8295 / 166491473 (0.004982%)|8674 / 30652 (28.298317%)|0.724180|0.721477|0.721440|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.950991|0.000000%|9.801848%|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|0.868094|0.000838%|26.380368%|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|0.867636|0.000852%|26.472040%|
</div>
@@ -86,9 +92,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.810903|0.871591|0.999963|0.935777|0.999939|0.000061|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.726732|0.717865|0.999950|0.858908|0.999898|0.000102|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.725993|0.717017|0.999950|0.858484|0.999898|0.000102|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.901982|1.000000|0.950991|0.999981|0.000019|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.945609|0.736196|0.999992|0.868094|0.999939|0.000061|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.944646|0.735280|0.999991|0.867636|0.999939|0.000061|
</details>
@@ -96,19 +102,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.822355|0.840152|0.858737|0.724364|0.840699|0.840669|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.724941|0.722271|0.719621|0.565278|0.722285|0.722234|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.724180|0.721477|0.718794|0.564305|0.721491|0.721440|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.840122|0.983845|0.986802|0.985321|0.985321|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.722221|0.980542|0.964998|0.972708|0.972708|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.721426|0.980461|0.964862|0.972599|0.972599|
|1|Radixor|PRIMARY_OUTPUT|0.978728|0.948465|0.920017|0.901982|0.949727|0.949718|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.894709|0.827865|0.770315|0.706288|0.834359|0.834331|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.893748|0.826916|0.769384|0.704908|0.833414|0.833386|
</details>
@@ -116,61 +112,39 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|26716|6230|3936|166485243|6230 / 166491473|3936 / 30652|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|22004|8274|8648|166483199|8274 / 166491473|8648 / 30652|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|21978|8295|8674|166483178|8295 / 166491473|8674 / 30652|
|1|Radixor|PRIMARY_OUTPUT|25582|0|2780|143394154|0 / 143394154|2780 / 28362|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|20880|1201|7482|143392953|1201 / 143394154|7482 / 28362|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|20854|1222|7508|143392932|1222 / 143394154|7508 / 28362|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 166491473 (0.000000%)|0 / 30652 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|30652|0|0|166491473|0 / 166491473|0 / 30652|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 143394154|0 / 28362|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999960|13214 / 166491473 (0.007937%)|0 / 30652 (0.000000%)|0.743562|0.822674|0.835888|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -178,7 +152,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.698764|1.000000|0.999921|0.999960|0.999921|0.000079|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -186,15 +160,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.743562|0.822674|0.920624|0.698764|0.835921|0.835888|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -202,7 +168,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|30652|13214|0|166478259|13214 / 166491473|0 / 30652|
|1|Radixor|ALL_CANDIDATES|28362|0|0|143394154|0 / 143394154|0 / 28362|
</details>
@@ -212,21 +178,21 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|3936|6230|6984|2404|13.172603%|5|21513|
|Radixor|2780|0|0|1091|6.441519%|5|18255|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.935853|6230 / 165926276 (0.003755%)|3924 / 30595 (12.825625%)|0.822169|0.840084|0.840609|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.859037|8274 / 165926276 (0.004987%)|8624 / 30595 (28.187612%)|0.724757|0.722255|0.722216|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.858661|8274 / 165926276 (0.004987%)|8647 / 30595 (28.262788%)|0.724438|0.721772|0.721734|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.951104|0.000000%|9.779191%|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|0.868252|0.000841%|26.348702%|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|0.867846|0.000841%|26.429959%|
</div>
@@ -234,9 +200,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.810644|0.871744|0.999962|0.935853|0.999939|0.000061|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.726434|0.718124|0.999950|0.859037|0.999898|0.000102|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.726226|0.717372|0.999950|0.858661|0.999898|0.000102|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.902208|1.000000|0.951104|0.999981|0.000019|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.945528|0.736513|0.999992|0.868252|0.999939|0.000061|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.945471|0.735700|0.999992|0.867846|0.999939|0.000061|
</details>
@@ -244,19 +210,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.822169|0.840084|0.858798|0.724263|0.840639|0.840609|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.724757|0.722255|0.719771|0.565258|0.722267|0.722216|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.724438|0.721772|0.719126|0.564666|0.721785|0.721734|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.840054|0.983815|0.986842|0.985326|0.985326|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.722204|0.980506|0.965065|0.972724|0.972724|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.721721|0.980506|0.964945|0.972663|0.972663|
|1|Radixor|PRIMARY_OUTPUT|0.978782|0.948590|0.920206|0.902208|0.949846|0.949837|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|0.894744|0.828034|0.770581|0.706534|0.834502|0.834474|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|0.894463|0.827499|0.769862|0.705755|0.834016|0.833989|
</details>
@@ -264,61 +220,39 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|26671|6230|3924|165920046|6230 / 165926276|3924 / 30595|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|21971|8274|8624|165918002|8274 / 165926276|8624 / 30595|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|21948|8274|8647|165918002|8274 / 165926276|8647 / 30595|
|1|Radixor|PRIMARY_OUTPUT|25537|0|2768|142869660|0 / 142869660|2768 / 28305|
|2|SNOWBALL NORWEGIAN NYNORSK DIRECT|PRIMARY_OUTPUT|20847|1201|7458|142868459|1201 / 142869660|7458 / 28305|
|3|SNOWBALL NORWEGIAN NYNORSK LUCENE FILTER|PRIMARY_OUTPUT|20824|1201|7481|142868459|1201 / 142869660|7481 / 28305|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 165926276 (0.000000%)|0 / 30595 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|30595|0|0|165926276|0 / 165926276|0 / 30595|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 142869660|0 / 28305|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999960|13214 / 165926276 (0.007964%)|0 / 30595 (0.000000%)|0.743207|0.822402|0.835654|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -326,7 +260,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.698372|1.000000|0.999920|0.999960|0.999920|0.000080|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -334,15 +268,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.743207|0.822402|0.920488|0.698372|0.835687|0.835654|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -350,7 +276,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|30595|13214|0|165913062|13214 / 165926276|0 / 30595|
|1|Radixor|ALL_CANDIDATES|28305|0|0|142869660|0 / 142869660|0 / 28305|
</details>
@@ -360,19 +286,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|3924|6230|6984|2399|13.167572%|5|21477|
|Radixor|2768|0|0|1086|6.423755%|5|18219|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -381,16 +307,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `NN_NO`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

249
docs/benchmarks/languages/persian.md
View File

@@ -8,13 +8,13 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `FA_IR` | 69 | 3,770 | 138 | 3,632 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `fa-ir-default` | `1.0.0` | `FA_IR` | 69 | 3,770 | 138 | 3,632 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **3,770**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **3,770**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
@@ -31,14 +31,20 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Radixor | 95.836% | 95.677% | 100.000% | Full Radixor dictionary patch-command stemmer. |
| Lucene PersianStemFilter | 1.485% | 0.000% | 40.580% | Lucene Persian suffix stemmer with required normalization in the measured path. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `persianRadixor` | 0.245 | 0.025 | 49.0 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene PersianStemFilter | `persianLucenePersianStemFilter` | 0.466 | 0.015 | 93.3 | 1.902 | Persian suffix stemmer with Lucene normalization in the measured path. |
| Radixor | `persianRadixor` | 0.243 | 0.004 | 66.9 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene PersianStemFilter | `persianLucenePersianStemFilter` | 0.469 | 0.007 | 129.1 | 1.930 | Persian suffix stemmer with Lucene normalization in the measured path. |
## Interpretation Notes
@@ -52,28 +58,28 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `FA_IR` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `FA_IR` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/fa_ir/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `fa-ir-default`, loaded from classpath resource `org/egothor/stemmer/models/fa-ir-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.974922** among 2 deterministic stemmers. The runner-up is `PERSIAN LUCENE PERSIAN STEM FILTER` at 0.502171, a difference of 0.472751. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.974922** among 2 deterministic stemmers. The runner-up is `PERSIAN LUCENE PERSIAN STEM FILTER` at 0.502171, a difference of 0.472751. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.976360** among 2 deterministic stemmers. The runner-up is `PERSIAN LUCENE PERSIAN STEM FILTER` at 0.502212, a difference of 0.474148. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.976360** among 2 deterministic stemmers. The runner-up is `PERSIAN LUCENE PERSIAN STEM FILTER` at 0.502212, a difference of 0.474148. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.974922|8621 / 6748402 (0.127749%)|4812 / 98448 (4.887860%)|0.922566|0.933071|0.932249|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.502171|179 / 6748402 (0.002652%)|98018 / 98448 (99.563221%)|0.021312|0.008682|0.054801|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.976360|0.000000%|4.728041%|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|0.502212|0.000049%|99.557494%|
</div>
@@ -81,8 +87,8 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.915693|0.951121|0.998723|0.974922|0.998038|0.001962|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.706076|0.004368|0.999973|0.502171|0.985658|0.014342|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.952720|1.000000|0.976360|0.999277|0.000723|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.992991|0.004425|1.000000|0.502212|0.984769|0.015231|
</details>
@@ -90,17 +96,8 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.922566|0.933071|0.943818|0.874539|0.933239|0.932249|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.021312|0.008682|0.005451|0.004360|0.055534|0.054801|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.932076|0.980343|0.984586|0.982460|0.982460|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.008507|0.985686|0.520347|0.681125|0.681125|
|1|Radixor|PRIMARY_OUTPUT|0.990172|0.975787|0.961815|0.952720|0.976074|0.975715|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.021738|0.008811|0.005525|0.004425|0.066288|0.065774|
</details>
@@ -108,60 +105,38 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|93636|8621|4812|6739781|8621 / 6748402|4812 / 98448|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|430|179|98018|6748223|179 / 6748402|98018 / 98448|
|1|Radixor|PRIMARY_OUTPUT|91503|0|4541|6182152|0 / 6182152|4541 / 96044|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|425|3|95619|6182149|3 / 6182152|95619 / 96044|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 6748402 (0.000000%)|0 / 98448 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|98448|0|0|6748402|0 / 6748402|0 / 98448|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 6182152|0 / 96044|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999005|13433 / 6748402 (0.199055%)|0 / 98448 (0.000000%)|0.901585|0.936133|0.937114|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -169,7 +144,7 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.879935|1.000000|0.998009|0.999005|0.998038|0.001962|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -177,15 +152,7 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.901585|0.936133|0.973435|0.879935|0.938048|0.937114|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -193,7 +160,7 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|98448|13433|0|6734969|13433 / 6748402|0 / 98448|
|1|Radixor|ALL_CANDIDATES|96044|0|0|6182152|0 / 6182152|0 / 96044|
</details>
@@ -203,20 +170,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|4812|8621|4812|314|8.484193%|2|4015|
|Radixor|4541|0|0|157|4.430023%|2|3701|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.974922|8621 / 6748402 (0.127749%)|4812 / 98448 (4.887860%)|0.922566|0.933071|0.932249|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.502171|179 / 6748402 (0.002652%)|98018 / 98448 (99.563221%)|0.021312|0.008682|0.054801|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.976360|0.000000%|4.728041%|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|0.502212|0.000049%|99.557494%|
</div>
@@ -224,8 +191,8 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.915693|0.951121|0.998723|0.974922|0.998038|0.001962|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.706076|0.004368|0.999973|0.502171|0.985658|0.014342|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.952720|1.000000|0.976360|0.999277|0.000723|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.992991|0.004425|1.000000|0.502212|0.984769|0.015231|
</details>
@@ -233,17 +200,8 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.922566|0.933071|0.943818|0.874539|0.933239|0.932249|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.021312|0.008682|0.005451|0.004360|0.055534|0.054801|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.932076|0.980343|0.984586|0.982460|0.982460|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.008507|0.985686|0.520347|0.681125|0.681125|
|1|Radixor|PRIMARY_OUTPUT|0.990172|0.975787|0.961815|0.952720|0.976074|0.975715|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|0.021738|0.008811|0.005525|0.004425|0.066288|0.065774|
</details>
@@ -251,60 +209,38 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|93636|8621|4812|6739781|8621 / 6748402|4812 / 98448|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|430|179|98018|6748223|179 / 6748402|98018 / 98448|
|1|Radixor|PRIMARY_OUTPUT|91503|0|4541|6182152|0 / 6182152|4541 / 96044|
|2|PERSIAN LUCENE PERSIAN STEM FILTER|PRIMARY_OUTPUT|425|3|95619|6182149|3 / 6182152|95619 / 96044|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 6748402 (0.000000%)|0 / 98448 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|98448|0|0|6748402|0 / 6748402|0 / 98448|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 6182152|0 / 96044|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999005|13433 / 6748402 (0.199055%)|0 / 98448 (0.000000%)|0.901585|0.936133|0.937114|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -312,7 +248,7 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.879935|1.000000|0.998009|0.999005|0.998038|0.001962|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -320,15 +256,7 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.901585|0.936133|0.973435|0.879935|0.938048|0.937114|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -336,7 +264,7 @@ This mode contains **4 result rows**, **2 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|98448|13433|0|6734969|13433 / 6748402|0 / 98448|
|1|Radixor|ALL_CANDIDATES|96044|0|0|6182152|0 / 6182152|0 / 96044|
</details>
@@ -346,19 +274,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|4812|8621|4812|314|8.484193%|2|4015|
|Radixor|4541|0|0|157|4.430023%|2|3701|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -367,16 +295,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `FA_IR`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

391
docs/benchmarks/languages/polish.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `PL_PL` | 9,990 | 132,308 | 19,957 | 112,351 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `pl-pl-unimorph` | `1.0.0` | `PL_PL` | 9,990 | 132,308 | 19,957 | 112,351 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **132,308**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **132,308**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 1,719 | 1.299% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 53,303 | 40.287% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 37,051 | 28.004% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 20,415 | 15.430% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 19,820 | 14.980% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 1,836 | 1.388% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 52,996 | 40.055% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 37,137 | 28.069% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 20,219 | 15.282% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 20,120 | 15.207% |
## Accuracy
@@ -36,17 +36,23 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene StempelFilter | 70.009% | 69.262% | 74.220% | Lucene TokenFilter integration path for table-driven Polish Stempel. |
| Lucene StempelStemmer direct | 70.009% | 69.262% | 74.220% | Direct table-driven Polish Stempel stemmer API. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `polishRadixor` | 9.049 | 0.485 | 80.5 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 483.316 | 11.455 | 4301.8 | 53.408 | Benchmark-only Polish Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene StempelStemmer direct | `polishLuceneStempelStemmerDirect` | 41.932 | 1.916 | 373.2 | 4.634 | Direct table-driven Polish Stempel stemmer API. |
| Lucene StempelFilter | `polishLuceneStempelFilter` | 45.277 | 13.693 | 403.0 | 5.003 | Lucene TokenFilter integration path for table-driven Polish Stempel. |
| Lucene MorfologikFilter | `polishLuceneMorfologikFilter` | 135.763 | 31.634 | 1208.4 | 15.002 | Dictionary-based Morfologik TokenFilter; may emit multiple terms. |
| Radixor | `polishRadixor` | 8.972 | 0.203 | 79.9 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 524.081 | 35.121 | 4664.7 | 58.412 | Benchmark-only Polish Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene StempelStemmer direct | `polishLuceneStempelStemmerDirect` | 37.947 | 0.335 | 337.8 | 4.229 | Direct table-driven Polish Stempel stemmer API. |
| Lucene StempelFilter | `polishLuceneStempelFilter` | 43.090 | 0.411 | 383.5 | 4.803 | Lucene TokenFilter integration path for table-driven Polish Stempel. |
| Lucene MorfologikFilter | `polishLuceneMorfologikFilter` | 143.527 | 1.176 | 1277.5 | 15.997 | Dictionary-based Morfologik TokenFilter; may emit multiple terms. |
## Interpretation Notes
@@ -60,31 +66,31 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `PL_PL` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `PL_PL` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/pl_pl/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `pl-pl-unimorph`, loaded from classpath resource `org/egothor/stemmer/models/pl-pl-unimorph/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.990388** among 5 deterministic stemmers. The runner-up is `POLISH LUCENE MORFOLOGIK FILTER` at 0.948154, a difference of 0.042234. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.990579** among 5 deterministic stemmers. The runner-up is `POLISH LUCENE MORFOLOGIK FILTER` at 0.948177, a difference of 0.042402. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.991105** among 5 deterministic stemmers. The runner-up is `POLISH LUCENE MORFOLOGIK FILTER` at 0.948392, a difference of 0.042713. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.991301** among 5 deterministic stemmers. The runner-up is `POLISH LUCENE MORFOLOGIK FILTER` at 0.948417, a difference of 0.042884. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.990388|13669 / 7482478003 (0.000183%)|21547 / 1120967 (1.922180%)|0.986324|0.984237|0.984241|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.948154|99228 / 7482478003 (0.001326%)|116220 / 1120967 (10.367834%)|0.907324|0.903167|0.903179|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.933222|52652 / 7482478003 (0.000704%)|149705 / 1120967 (13.354987%)|0.930930|0.905656|0.906571|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.855748|66669 / 7482478003 (0.000891%)|323394 / 1120967 (28.849556%)|0.871106|0.803515|0.810296|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.855748|66669 / 7482478003 (0.000891%)|323394 / 1120967 (28.849556%)|0.871106|0.803515|0.810296|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.991105|0.000000%|1.779024%|
|2|POLISH LUCENE MORFOLOGIK FILTER|0.948392|0.001042%|10.320543%|
|3|HUNSPELL POLISH LUCENE FILTER|0.933457|0.000383%|13.308172%|
|4|POLISH LUCENE STEMPEL DIRECT|0.855699|0.000602%|28.859618%|
|5|POLISH LUCENE STEMPEL FILTER|0.855699|0.000602%|28.859618%|
</div>
@@ -92,11 +98,11 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987720|0.980778|0.999998|0.990388|0.999995|0.000005|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.910118|0.896322|0.999987|0.948154|0.999971|0.000029|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.948578|0.866450|0.999993|0.933222|0.999973|0.000027|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.922858|0.711504|0.999991|0.855748|0.999948|0.000052|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.922858|0.711504|0.999991|0.855748|0.999948|0.000052|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.982210|1.000000|0.991105|0.999997|0.000003|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.929398|0.896795|0.999990|0.948392|0.999974|0.000026|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.971931|0.866918|0.999996|0.933457|0.999976|0.000024|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.947549|0.711404|0.999994|0.855699|0.999950|0.000050|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.947549|0.711404|0.999994|0.855699|0.999950|0.000050|
</details>
@@ -104,23 +110,11 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986324|0.984237|0.982159|0.968963|0.984243|0.984241|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.907324|0.903167|0.899047|0.823432|0.903193|0.903179|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.930930|0.905656|0.881718|0.827579|0.906584|0.906571|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.871106|0.803515|0.745659|0.671564|0.810320|0.810296|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.871106|0.803515|0.745659|0.671564|0.810320|0.810296|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.984234|0.996967|0.996469|0.996718|0.996718|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.903153|0.990022|0.977054|0.983495|0.983495|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.905642|0.994546|0.970520|0.982386|0.982386|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.803490|0.991767|0.931069|0.960460|0.960460|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.803490|0.991767|0.931069|0.960460|0.960460|
|1|Radixor|PRIMARY_OUTPUT|0.996391|0.991025|0.985717|0.982210|0.991065|0.991064|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.922689|0.912805|0.903131|0.839597|0.912951|0.912938|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.948942|0.916426|0.886065|0.845744|0.917924|0.917913|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.888559|0.812669|0.748723|0.684450|0.821030|0.821007|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.888559|0.812669|0.748723|0.684450|0.821030|0.821007|
</details>
@@ -128,75 +122,47 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|1099420|13669|21547|7482464334|13669 / 7482478003|21547 / 1120967|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|1004747|99228|116220|7482378775|99228 / 7482478003|116220 / 1120967|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|971262|52652|149705|7482425351|52652 / 7482478003|149705 / 1120967|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|797573|66669|323394|7482411334|66669 / 7482478003|323394 / 1120967|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|797573|66669|323394|7482411334|66669 / 7482478003|323394 / 1120967|
|1|Radixor|PRIMARY_OUTPUT|1097200|0|19873|7303238338|0 / 7303238338|19873 / 1117073|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|1001785|76101|115288|7303162237|76101 / 7303238338|115288 / 1117073|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|968411|27967|148662|7303210371|27967 / 7303238338|148662 / 1117073|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|794690|43990|322383|7303194348|43990 / 7303238338|322383 / 1117073|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|794690|43990|322383|7303194348|43990 / 7303238338|322383 / 1117073|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 7482478003 (0.000000%)|0 / 1120967 (0.000000%)|1.000000|1.000000|1.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.987570|85532 / 7482478003 (0.001143%)|27855 / 1120967 (2.484908%)|0.936598|0.950693|0.950985|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|0.963982|42213 / 7482478003 (0.000564%)|80743 / 1120967 (7.202977%)|0.954209|0.944197|0.944333|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL POLISH LUCENE FILTER|0.000356%|7.227639%|
|POLISH LUCENE MORFOLOGIK FILTER|0.001000%|2.493123%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.927432|0.975151|0.999989|0.987570|0.999985|0.000015|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|0.961002|0.927970|0.999994|0.963982|0.999984|0.000016|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.936598|0.950693|0.965218|0.906020|0.950992|0.950985|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|0.954209|0.944197|0.934394|0.894293|0.944342|0.944333|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1120967|0|0|7482478003|0 / 7482478003|0 / 1120967|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|1093112|85532|27855|7482392471|85532 / 7482478003|27855 / 1120967|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|1040224|42213|80743|7482435790|42213 / 7482478003|80743 / 1120967|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL POLISH LUCENE FILTER|25967 / 7303238338|80738 / 1117073|
|POLISH LUCENE MORFOLOGIK FILTER|73019 / 7303238338|27850 / 1117073|
|Radixor|0 / 7303238338|0 / 1117073|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999997|38073 / 7482478003 (0.000509%)|0 / 1120967 (0.000000%)|0.973547|0.983301|0.983436|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.987566|143096 / 7482478003 (0.001912%)|27855 / 1120967 (2.484908%)|0.901045|0.927476|0.928576|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.963980|82745 / 7482478003 (0.001106%)|80743 / 1120967 (7.202977%)|0.926646|0.927142|0.927132|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|POLISH LUCENE MORFOLOGIK FILTER|0.987528|0.001376%|2.493123%|
|3|HUNSPELL POLISH LUCENE FILTER|0.963859|0.000609%|7.227639%|
</div>
@@ -204,9 +170,9 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.967151|1.000000|0.999995|0.999997|0.999995|0.000005|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.884246|0.975151|0.999981|0.987566|0.999977|0.000023|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.926316|0.927970|0.999989|0.963980|0.999978|0.000022|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.915516|0.975069|0.999986|0.987528|0.999982|0.000018|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.958830|0.927724|0.999994|0.963859|0.999983|0.000017|
</details>
@@ -214,19 +180,9 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.973547|0.983301|0.993253|0.967151|0.983438|0.983436|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.901045|0.927476|0.955505|0.864761|0.928587|0.928576|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.926646|0.927142|0.927639|0.864180|0.927143|0.927132|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.926837|0.944354|0.962546|0.894575|0.944823|0.944815|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.952443|0.943020|0.933782|0.892184|0.943149|0.943140|
</details>
@@ -234,9 +190,9 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1120967|38073|0|7482439930|38073 / 7482478003|0 / 1120967|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|1093112|143096|27855|7482334907|143096 / 7482478003|27855 / 1120967|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|1040224|82745|80743|7482395258|82745 / 7482478003|80743 / 1120967|
|1|Radixor|ALL_CANDIDATES|1117073|0|0|7303238338|0 / 7303238338|0 / 1117073|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|1089223|100514|27850|7303137824|100514 / 7303238338|27850 / 1117073|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|1036335|44498|80738|7303193840|44498 / 7303238338|80738 / 1117073|
</details>
@@ -246,25 +202,25 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL POLISH LUCENE FILTER|68962|10439|30093|11447|9.356634%|6|135231|
|POLISH LUCENE MORFOLOGIK FILTER|88365|13696|43868|12873|10.522229%|5|136636|
|Radixor|21547|13669|24404|2866|2.342632%|4|125778|
|HUNSPELL POLISH LUCENE FILTER|67924|2000|16531|10485|8.674824%|6|132492|
|POLISH LUCENE MORFOLOGIK FILTER|87438|3082|24413|11776|9.742941%|5|133810|
|Radixor|19873|0|0|1392|1.151679%|4|122430|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.990579|13669 / 7310252699 (0.000187%)|21000 / 1114651 (1.883998%)|0.986350|0.984397|0.984400|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.948177|99224 / 7310252699 (0.001357%)|115513 / 1114651 (10.363154%)|0.906972|0.902966|0.902976|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.933309|51950 / 7310252699 (0.000711%)|148667 / 1114651 (13.337538%)|0.931269|0.905928|0.906847|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.856382|66274 / 7310252699 (0.000907%)|320158 / 1114651 (28.722712%)|0.871591|0.804380|0.811082|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.856382|66274 / 7310252699 (0.000907%)|320158 / 1114651 (28.722712%)|0.871591|0.804380|0.811082|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.991301|0.000000%|1.739895%|
|2|POLISH LUCENE MORFOLOGIK FILTER|0.948417|0.001067%|10.315578%|
|3|HUNSPELL POLISH LUCENE FILTER|0.933546|0.000382%|13.290396%|
|4|POLISH LUCENE STEMPEL DIRECT|0.856335|0.000611%|28.732387%|
|5|POLISH LUCENE STEMPEL FILTER|0.856335|0.000611%|28.732387%|
</div>
@@ -272,11 +228,11 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987656|0.981160|0.999998|0.990579|0.999995|0.000005|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.909662|0.896368|0.999986|0.948177|0.999971|0.000029|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.948965|0.866625|0.999993|0.933309|0.999973|0.000027|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.923006|0.712773|0.999991|0.856382|0.999947|0.000053|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.923006|0.712773|0.999991|0.856382|0.999947|0.000053|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.982601|1.000000|0.991301|0.999997|0.000003|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.929032|0.896844|0.999989|0.948417|0.999973|0.000027|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.972469|0.867096|0.999996|0.933546|0.999975|0.000025|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.947796|0.712676|0.999994|0.856335|0.999949|0.000051|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.947796|0.712676|0.999994|0.856335|0.999949|0.000051|
</details>
@@ -284,23 +240,11 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986350|0.984397|0.982452|0.969274|0.984403|0.984400|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.906972|0.902966|0.898996|0.823098|0.902991|0.902976|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.931269|0.905928|0.881929|0.828033|0.906861|0.906847|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.871591|0.804380|0.746792|0.672772|0.811106|0.811082|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.871591|0.804380|0.746792|0.672772|0.811106|0.811082|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.984395|0.996926|0.996647|0.996786|0.996786|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.902952|0.989889|0.977012|0.983408|0.983408|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.905914|0.994584|0.970514|0.982402|0.982402|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.804354|0.991711|0.931318|0.960566|0.960566|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.804354|0.991711|0.931318|0.960566|0.960566|
|1|Radixor|PRIMARY_OUTPUT|0.996471|0.991224|0.986032|0.982601|0.991262|0.991261|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.922411|0.912654|0.903102|0.839342|0.912796|0.912783|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|0.949394|0.916764|0.886303|0.846320|0.918272|0.918260|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|0.889130|0.813590|0.749881|0.685758|0.821871|0.821848|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|0.889130|0.813590|0.749881|0.685758|0.821871|0.821848|
</details>
@@ -308,75 +252,47 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|1093651|13669|21000|7310239030|13669 / 7310252699|21000 / 1114651|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|999138|99224|115513|7310153475|99224 / 7310252699|115513 / 1114651|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|965984|51950|148667|7310200749|51950 / 7310252699|148667 / 1114651|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|794493|66274|320158|7310186425|66274 / 7310252699|320158 / 1114651|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|794493|66274|320158|7310186425|66274 / 7310252699|320158 / 1114651|
|1|Radixor|PRIMARY_OUTPUT|1091431|0|19326|7133100218|0 / 7133100218|19326 / 1110757|
|2|POLISH LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|996176|76097|114581|7133024121|76097 / 7133100218|114581 / 1110757|
|3|HUNSPELL POLISH LUCENE FILTER|PRIMARY_OUTPUT|963133|27267|147624|7133072951|27267 / 7133100218|147624 / 1110757|
|4|POLISH LUCENE STEMPEL DIRECT|PRIMARY_OUTPUT|791610|43601|319147|7133056617|43601 / 7133100218|319147 / 1110757|
|5|POLISH LUCENE STEMPEL FILTER|PRIMARY_OUTPUT|791610|43601|319147|7133056617|43601 / 7133100218|319147 / 1110757|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 7310252699 (0.000000%)|0 / 1114651 (0.000000%)|1.000000|1.000000|1.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.987661|85532 / 7310252699 (0.001170%)|27494 / 1114651 (2.466602%)|0.936331|0.950586|0.950885|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|0.963946|41671 / 7310252699 (0.000570%)|80368 / 1114651 (7.210149%)|0.954406|0.944290|0.944429|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL POLISH LUCENE FILTER|0.000356%|7.234976%|
|POLISH LUCENE MORFOLOGIK FILTER|0.001024%|2.474799%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.927063|0.975334|0.999988|0.987661|0.999985|0.000015|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|0.961271|0.927899|0.999994|0.963946|0.999983|0.000017|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.936331|0.950586|0.965282|0.905826|0.950892|0.950885|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|0.954406|0.944290|0.934386|0.894459|0.944437|0.944429|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1114651|0|0|7310252699|0 / 7310252699|0 / 1114651|
|2|POLISH LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|1087157|85532|27494|7310167167|85532 / 7310252699|27494 / 1114651|
|3|HUNSPELL POLISH LUCENE FILTER|ANY_CANDIDATE|1034283|41671|80368|7310211028|41671 / 7310252699|80368 / 1114651|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL POLISH LUCENE FILTER|25425 / 7133100218|80363 / 1110757|
|POLISH LUCENE MORFOLOGIK FILTER|73019 / 7133100218|27489 / 1110757|
|Radixor|0 / 7133100218|0 / 1110757|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999997|38073 / 7310252699 (0.000521%)|0 / 1114651 (0.000000%)|0.973401|0.983208|0.983344|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.987657|143085 / 7310252699 (0.001957%)|27494 / 1114651 (2.466602%)|0.900618|0.927255|0.928372|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.963944|81865 / 7310252699 (0.001120%)|80368 / 1114651 (7.210149%)|0.926903|0.927276|0.927265|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|POLISH LUCENE MORFOLOGIK FILTER|0.987619|0.001409%|2.474799%|
|3|HUNSPELL POLISH LUCENE FILTER|0.963822|0.000612%|7.234976%|
</div>
@@ -384,9 +300,9 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.966971|1.000000|0.999995|0.999997|0.999995|0.000005|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.883694|0.975334|0.999980|0.987657|0.999977|0.000023|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.926654|0.927899|0.999989|0.963944|0.999978|0.000022|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.915099|0.975252|0.999986|0.987619|0.999982|0.000018|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.959377|0.927650|0.999994|0.963822|0.999983|0.000017|
</details>
@@ -394,19 +310,9 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.973401|0.983208|0.993215|0.966971|0.983347|0.983344|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.900618|0.927255|0.955516|0.864376|0.928384|0.928372|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.926903|0.927276|0.927649|0.864412|0.927276|0.927265|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.926529|0.944219|0.962597|0.894332|0.944697|0.944688|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|0.952859|0.943247|0.933827|0.892590|0.943380|0.943372|
</details>
@@ -414,9 +320,9 @@ This mode contains **11 result rows**, **5 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|1114651|38073|0|7310214626|38073 / 7310252699|0 / 1114651|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|1087157|143085|27494|7310109614|143085 / 7310252699|27494 / 1114651|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|1034283|81865|80368|7310170834|81865 / 7310252699|80368 / 1114651|
|1|Radixor|ALL_CANDIDATES|1110757|0|0|7133100218|0 / 7133100218|0 / 1110757|
|2|POLISH LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|1083268|100503|27489|7132999715|100503 / 7133100218|27489 / 1110757|
|3|HUNSPELL POLISH LUCENE FILTER|ALL_CANDIDATES|1030394|43630|80363|7133056588|43630 / 7133100218|80363 / 1110757|
</details>
@@ -426,21 +332,21 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL POLISH LUCENE FILTER|68299|10279|29915|11265|9.315692%|6|133595|
|POLISH LUCENE MORFOLOGIK FILTER|88019|13692|43861|12763|10.554476%|5|135105|
|Radixor|21000|13669|24404|2780|2.298946%|4|124274|
|HUNSPELL POLISH LUCENE FILTER|67261|1842|16363|10303|8.625294%|6|130856|
|POLISH LUCENE MORFOLOGIK FILTER|87092|3078|24406|11666|9.766348%|5|132279|
|Radixor|19326|0|0|1306|1.093335%|4|120926|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -449,16 +355,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `PL_PL`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

333
docs/benchmarks/languages/portuguese.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `PT_PT` | 4,001 | 215,490 | 8,002 | 207,488 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `pt-pt-default` | `1.0.0` | `PT_PT` | 4,001 | 215,490 | 8,002 | 207,488 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **215,490**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **215,490**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 3,806 | 1.766% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 120,691 | 56.008% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 120,535 | 55.935% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 71,284 | 33.080% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 8,003 | 3.714% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 11,706 | 5.432% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 11,862 | 5.505% |
## Accuracy
@@ -37,18 +37,24 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 0.625% | 0.558% | 2.374% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene PortugueseStemFilter | 0.312% | 0.308% | 0.425% | Portuguese RSLP-style Lucene TokenFilter stemmer. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `portugueseRadixor` | 12.109 | 0.698 | 58.4 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene PortugueseLightStemFilter | `portugueseLucenePortugueseLightStemFilter` | 11.172 | 1.870 | 53.8 | 0.923 | Light Portuguese suffix stemmer. |
| Lucene PortugueseMinimalStemFilter | `portugueseLucenePortugueseMinimalStemFilter` | 16.038 | 1.752 | 77.3 | 1.325 | Minimal Portuguese suffix reducer. |
| Official Snowball direct | `snowballDirect[PORTUGUESE]` | 53.725 | 5.356 | 258.9 | 4.437 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[PORTUGUESE]` | 57.457 | 1.182 | 276.9 | 4.745 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Lucene PortugueseStemFilter | `portugueseLucenePortugueseStemFilter` | 165.447 | 40.334 | 797.4 | 13.663 | Portuguese RSLP-style Lucene TokenFilter. |
| Radixor | `portugueseRadixor` | 12.301 | 0.252 | 59.3 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene PortugueseLightStemFilter | `portugueseLucenePortugueseLightStemFilter` | 11.409 | 0.151 | 55.0 | 0.927 | Light Portuguese suffix stemmer. |
| Lucene PortugueseMinimalStemFilter | `portugueseLucenePortugueseMinimalStemFilter` | 15.619 | 0.084 | 75.3 | 1.270 | Minimal Portuguese suffix reducer. |
| Official Snowball direct | `snowballDirect[PORTUGUESE]` | 57.577 | 1.591 | 277.5 | 4.681 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[PORTUGUESE]` | 63.403 | 2.720 | 305.6 | 5.154 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Lucene PortugueseStemFilter | `portugueseLucenePortugueseStemFilter` | 158.014 | 5.150 | 761.6 | 12.845 | Portuguese RSLP-style Lucene TokenFilter. |
## Interpretation Notes
@@ -62,32 +68,32 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `PT_PT` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `PT_PT` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/pt_pt/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `pt-pt-default`, loaded from classpath resource `org/egothor/stemmer/models/pt-pt-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.998502** among 6 deterministic stemmers. The runner-up is `SNOWBALL PORTUGUESE DIRECT` at 0.938800, a difference of 0.059702. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.998502** among 6 deterministic stemmers. The runner-up is `SNOWBALL PORTUGUESE DIRECT` at 0.938800, a difference of 0.059702. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.998542** among 6 deterministic stemmers. The runner-up is `SNOWBALL PORTUGUESE DIRECT` at 0.938922, a difference of 0.059620. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.998542** among 6 deterministic stemmers. The runner-up is `SNOWBALL PORTUGUESE DIRECT` at 0.938922, a difference of 0.059620. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.998502|20678 / 22358203756 (0.000092%)|16444 / 5489060 (0.299578%)|0.996389|0.996620|0.996619|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.938800|167230 / 22358203756 (0.000748%)|671821 / 5489060 (12.239272%)|0.947271|0.919888|0.920940|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.938800|167230 / 22358203756 (0.000748%)|671821 / 5489060 (12.239272%)|0.947271|0.919888|0.920940|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.846459|99075 / 22358203756 (0.000443%)|1685572 / 5489060 (30.707844%)|0.901330|0.809975|0.821750|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.513648|2511 / 22358203756 (0.000011%)|5339230 / 5489060 (97.270389%)|0.122843|0.053118|0.163828|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.503954|598 / 22358203756 (0.000003%)|5445654 / 5489060 (99.209227%)|0.038310|0.015690|0.088308|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.998542|0.000000%|0.291615%|
|2|SNOWBALL PORTUGUESE DIRECT|0.938922|0.000656%|12.214929%|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|0.938922|0.000656%|12.214929%|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|0.846554|0.000364%|30.688771%|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|0.513632|0.000006%|97.273598%|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|0.503949|&lt;0.000001%|99.210240%|
</div>
@@ -95,12 +101,12 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996236|0.997004|0.999999|0.998502|0.999998|0.000002|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.966450|0.877607|0.999993|0.938800|0.999962|0.000038|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.966450|0.877607|0.999993|0.938800|0.999962|0.000038|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.974613|0.692922|0.999996|0.846459|0.999920|0.000080|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.983517|0.027296|1.000000|0.513648|0.999761|0.000239|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.986410|0.007908|1.000000|0.503954|0.999756|0.000244|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.997084|1.000000|0.998542|0.999999|0.000001|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.970538|0.877851|0.999993|0.938922|0.999963|0.000037|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.970538|0.877851|0.999993|0.938922|0.999963|0.000037|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.979145|0.693112|0.999996|0.846554|0.999921|0.000079|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.991719|0.027264|1.000000|0.513632|0.999760|0.000240|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.999608|0.007898|1.000000|0.503949|0.999756|0.000244|
</details>
@@ -108,25 +114,12 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996389|0.996620|0.996850|0.993262|0.996620|0.996619|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.947271|0.919888|0.894045|0.851661|0.920958|0.920940|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.947271|0.919888|0.894045|0.851661|0.920958|0.920940|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.901330|0.809975|0.735434|0.680636|0.821785|0.821750|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.122843|0.053118|0.033885|0.027284|0.163848|0.163828|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.038310|0.015690|0.009865|0.007907|0.088319|0.088308|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996619|0.999299|0.999347|0.999323|0.999323|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.919870|0.996663|0.967924|0.982083|0.982083|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.919870|0.996663|0.967924|0.982083|0.982083|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.809936|0.996729|0.918475|0.956003|0.956003|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.053105|0.999226|0.720580|0.837330|0.837330|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.015686|0.999664|0.692383|0.818122|0.818122|
|1|Radixor|PRIMARY_OUTPUT|0.999415|0.998540|0.997666|0.997084|0.998541|0.998541|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.950467|0.921871|0.894944|0.855065|0.923032|0.923014|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.950467|0.921871|0.894944|0.855065|0.923032|0.923014|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.904492|0.811666|0.736120|0.683029|0.823807|0.823773|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.122815|0.053069|0.033847|0.027258|0.164433|0.164413|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.038278|0.015671|0.009853|0.007898|0.088851|0.088840|
</details>
@@ -134,64 +127,42 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|5472616|20678|16444|22358183078|20678 / 22358203756|16444 / 5489060|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|4817239|167230|671821|22358036526|167230 / 22358203756|671821 / 5489060|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|4817239|167230|671821|22358036526|167230 / 22358203756|671821 / 5489060|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|3803488|99075|1685572|22358104681|99075 / 22358203756|1685572 / 5489060|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|149830|2511|5339230|22358201245|2511 / 22358203756|5339230 / 5489060|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|43406|598|5445654|22358203158|598 / 22358203756|5445654 / 5489060|
|1|Radixor|PRIMARY_OUTPUT|5470353|0|15999|22274113243|0 / 22274113243|15999 / 5486352|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|4816198|146201|670154|22273967042|146201 / 22274113243|670154 / 5486352|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|4816198|146201|670154|22273967042|146201 / 22274113243|670154 / 5486352|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|3802658|80995|1683694|22274032248|80995 / 22274113243|1683694 / 5486352|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|149580|1249|5336772|22274111994|1249 / 22274113243|5336772 / 5486352|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|43329|17|5443023|22274113226|17 / 22274113243|5443023 / 5486352|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 22358203756 (0.000000%)|0 / 5489060 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|5489060|0|0|22358203756|0 / 22358203756|0 / 5489060|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 22274113243|0 / 5486352|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|38310 / 22358203756 (0.000171%)|0 / 5489060 (0.000000%)|0.994448|0.996522|0.996528|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -199,7 +170,7 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.993069|1.000000|0.999998|0.999999|0.999998|0.000002|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -207,15 +178,7 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.994448|0.996522|0.998606|0.993069|0.996528|0.996528|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -223,7 +186,7 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|5489060|38310|0|22358165446|38310 / 22358203756|0 / 5489060|
|1|Radixor|ALL_CANDIDATES|5486352|0|0|22274113243|0 / 22274113243|0 / 5486352|
</details>
@@ -233,24 +196,24 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|16444|20678|17632|790|0.373542%|3|212297|
|Radixor|15999|0|0|392|0.185702%|3|211489|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.998502|20678 / 22358203756 (0.000092%)|16444 / 5489060 (0.299578%)|0.996389|0.996620|0.996619|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.938800|167230 / 22358203756 (0.000748%)|671821 / 5489060 (12.239272%)|0.947271|0.919888|0.920940|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.938800|167230 / 22358203756 (0.000748%)|671821 / 5489060 (12.239272%)|0.947271|0.919888|0.920940|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.846459|99075 / 22358203756 (0.000443%)|1685572 / 5489060 (30.707844%)|0.901330|0.809975|0.821750|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.513648|2511 / 22358203756 (0.000011%)|5339230 / 5489060 (97.270389%)|0.122843|0.053118|0.163828|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.503954|598 / 22358203756 (0.000003%)|5445654 / 5489060 (99.209227%)|0.038310|0.015690|0.088308|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.998542|0.000000%|0.291615%|
|2|SNOWBALL PORTUGUESE DIRECT|0.938922|0.000656%|12.214929%|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|0.938922|0.000656%|12.214929%|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|0.846554|0.000364%|30.688771%|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|0.513632|0.000006%|97.273598%|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|0.503949|&lt;0.000001%|99.210240%|
</div>
@@ -258,12 +221,12 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996236|0.997004|0.999999|0.998502|0.999998|0.000002|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.966450|0.877607|0.999993|0.938800|0.999962|0.000038|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.966450|0.877607|0.999993|0.938800|0.999962|0.000038|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.974613|0.692922|0.999996|0.846459|0.999920|0.000080|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.983517|0.027296|1.000000|0.513648|0.999761|0.000239|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.986410|0.007908|1.000000|0.503954|0.999756|0.000244|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.997084|1.000000|0.998542|0.999999|0.000001|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.970538|0.877851|0.999993|0.938922|0.999963|0.000037|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.970538|0.877851|0.999993|0.938922|0.999963|0.000037|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.979145|0.693112|0.999996|0.846554|0.999921|0.000079|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.991719|0.027264|1.000000|0.513632|0.999760|0.000240|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.999608|0.007898|1.000000|0.503949|0.999756|0.000244|
</details>
@@ -271,25 +234,12 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996389|0.996620|0.996850|0.993262|0.996620|0.996619|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.947271|0.919888|0.894045|0.851661|0.920958|0.920940|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.947271|0.919888|0.894045|0.851661|0.920958|0.920940|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.901330|0.809975|0.735434|0.680636|0.821785|0.821750|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.122843|0.053118|0.033885|0.027284|0.163848|0.163828|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.038310|0.015690|0.009865|0.007907|0.088319|0.088308|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.996619|0.999299|0.999347|0.999323|0.999323|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.919870|0.996663|0.967924|0.982083|0.982083|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.919870|0.996663|0.967924|0.982083|0.982083|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.809936|0.996729|0.918475|0.956003|0.956003|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.053105|0.999226|0.720580|0.837330|0.837330|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.015686|0.999664|0.692383|0.818122|0.818122|
|1|Radixor|PRIMARY_OUTPUT|0.999415|0.998540|0.997666|0.997084|0.998541|0.998541|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|0.950467|0.921871|0.894944|0.855065|0.923032|0.923014|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|0.950467|0.921871|0.894944|0.855065|0.923032|0.923014|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|0.904492|0.811666|0.736120|0.683029|0.823807|0.823773|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|0.122815|0.053069|0.033847|0.027258|0.164433|0.164413|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.038278|0.015671|0.009853|0.007898|0.088851|0.088840|
</details>
@@ -297,64 +247,42 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|5472616|20678|16444|22358183078|20678 / 22358203756|16444 / 5489060|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|4817239|167230|671821|22358036526|167230 / 22358203756|671821 / 5489060|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|4817239|167230|671821|22358036526|167230 / 22358203756|671821 / 5489060|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|3803488|99075|1685572|22358104681|99075 / 22358203756|1685572 / 5489060|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|149830|2511|5339230|22358201245|2511 / 22358203756|5339230 / 5489060|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|43406|598|5445654|22358203158|598 / 22358203756|5445654 / 5489060|
|1|Radixor|PRIMARY_OUTPUT|5470353|0|15999|22274113243|0 / 22274113243|15999 / 5486352|
|2|SNOWBALL PORTUGUESE DIRECT|PRIMARY_OUTPUT|4816198|146201|670154|22273967042|146201 / 22274113243|670154 / 5486352|
|3|SNOWBALL PORTUGUESE LUCENE FILTER|PRIMARY_OUTPUT|4816198|146201|670154|22273967042|146201 / 22274113243|670154 / 5486352|
|4|PORTUGUESE LUCENE PORTUGUESE STEM FILTER|PRIMARY_OUTPUT|3802658|80995|1683694|22274032248|80995 / 22274113243|1683694 / 5486352|
|5|PORTUGUESE LUCENE PORTUGUESE LIGHT STEM FILTER|PRIMARY_OUTPUT|149580|1249|5336772|22274111994|1249 / 22274113243|5336772 / 5486352|
|6|PORTUGUESE LUCENE PORTUGUESE MINIMAL STEM FILTER|PRIMARY_OUTPUT|43329|17|5443023|22274113226|17 / 22274113243|5443023 / 5486352|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 22358203756 (0.000000%)|0 / 5489060 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|5489060|0|0|22358203756|0 / 22358203756|0 / 5489060|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 22274113243|0 / 5486352|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|38310 / 22358203756 (0.000171%)|0 / 5489060 (0.000000%)|0.994448|0.996522|0.996528|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -362,7 +290,7 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.993069|1.000000|0.999998|0.999999|0.999998|0.000002|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -370,15 +298,7 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.994448|0.996522|0.998606|0.993069|0.996528|0.996528|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -386,7 +306,7 @@ This mode contains **8 result rows**, **6 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|5489060|38310|0|22358165446|38310 / 22358203756|0 / 5489060|
|1|Radixor|ALL_CANDIDATES|5486352|0|0|22274113243|0 / 22274113243|0 / 5486352|
</details>
@@ -396,19 +316,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|16444|20678|17632|790|0.373542%|3|212297|
|Radixor|15999|0|0|392|0.185702%|3|211489|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -417,16 +337,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `PT_PT`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

299
docs/benchmarks/languages/russian.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `RU_RU` | 37,410 | 806,279 | 74,808 | 731,471 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `ru-ru-default` | `1.0.0` | `RU_RU` | 37,410 | 806,279 | 74,808 | 731,471 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **806,279**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **806,279**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 9,260 | 1.148% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 584,785 | 72.529% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 82,864 | 10.277% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 75,646 | 9.382% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 53,724 | 6.663% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 9,287 | 1.152% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 580,915 | 72.049% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 82,956 | 10.289% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 75,527 | 9.367% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 57,594 | 7.143% |
## Accuracy
@@ -35,16 +35,22 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene SnowballFilter | 9.162% | 8.162% | 18.936% | Lucene TokenFilter integration path around the Snowball algorithm. |
| Official Snowball direct | 9.162% | 8.162% | 18.936% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `russianRadixor` | 89.671 | 3.886 | 122.6 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene RussianLightStemFilter | `russianLuceneRussianLightStemFilter` | 60.522 | 5.310 | 82.7 | 0.675 | Light Russian suffix stemmer. |
| Official Snowball direct | `snowballDirect[RUSSIAN]` | 106.031 | 9.287 | 145.0 | 1.182 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[RUSSIAN]` | 137.512 | 10.801 | 188.0 | 1.534 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `russianRadixor` | 90.151 | 1.796 | 123.2 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene RussianLightStemFilter | `russianLuceneRussianLightStemFilter` | 59.456 | 2.102 | 81.3 | 0.660 | Light Russian suffix stemmer. |
| Official Snowball direct | `snowballDirect[RUSSIAN]` | 102.353 | 1.669 | 139.9 | 1.135 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[RUSSIAN]` | 138.597 | 4.727 | 189.5 | 1.537 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -58,30 +64,30 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `RU_RU` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `RU_RU` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/ru_ru/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `ru-ru-default`, loaded from classpath resource `org/egothor/stemmer/models/ru-ru-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.989827** among 4 deterministic stemmers. The runner-up is `SNOWBALL RUSSIAN LUCENE FILTER` at 0.834876, a difference of 0.154951. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.989852** among 4 deterministic stemmers. The runner-up is `SNOWBALL RUSSIAN DIRECT` at 0.834854, a difference of 0.154998. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.990188** among 4 deterministic stemmers. The runner-up is `SNOWBALL RUSSIAN LUCENE FILTER` at 0.834565, a difference of 0.155624. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.990213** among 4 deterministic stemmers. The runner-up is `SNOWBALL RUSSIAN DIRECT` at 0.834542, a difference of 0.155670. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989827|155850 / 295576291016 (0.000053%)|266302 / 13089505 (2.034470%)|0.986313|0.983806|0.983814|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.834876|3785790 / 295576291016 (0.001281%)|4322616 / 13089505 (33.023525%)|0.692485|0.683786|0.683923|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.834867|3782908 / 295576291016 (0.001280%)|4322849 / 13089505 (33.025305%)|0.692603|0.683851|0.683989|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.617692|321183 / 295576291016 (0.000109%)|10008438 / 13089505 (76.461547%)|0.577011|0.373649|0.461687|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.990188|0.000000%|1.962362%|
|2|SNOWBALL RUSSIAN LUCENE FILTER|0.834565|0.001215%|33.085867%|
|3|SNOWBALL RUSSIAN DIRECT|0.834556|0.001214%|33.087654%|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|0.616440|0.000059%|76.711890%|
</div>
@@ -89,10 +95,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987992|0.979655|0.999999|0.989827|0.999999|0.000001|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.698408|0.669765|0.999987|0.834876|0.999973|0.000027|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.698563|0.669747|0.999987|0.834867|0.999973|0.000027|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.905597|0.235385|0.999999|0.617692|0.999965|0.000035|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.980376|1.000000|0.990188|0.999999|0.000001|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.713518|0.669141|0.999988|0.834565|0.999973|0.000027|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.713680|0.669123|0.999988|0.834556|0.999973|0.000027|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.946958|0.232881|0.999999|0.616440|0.999965|0.000035|
</details>
@@ -100,21 +106,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986313|0.983806|0.981311|0.968128|0.983815|0.983814|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.692485|0.683786|0.675304|0.519510|0.683936|0.683923|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.692603|0.683851|0.675318|0.519585|0.684003|0.683989|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.577011|0.373649|0.276278|0.229747|0.461696|0.461687|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.983805|0.997699|0.997274|0.997487|0.997487|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.683773|0.974131|0.953674|0.963794|0.963794|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.683838|0.974180|0.953661|0.963811|0.963811|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.373638|0.994311|0.870888|0.928516|0.928516|
|1|Radixor|PRIMARY_OUTPUT|0.996013|0.990091|0.984239|0.980376|0.990140|0.990139|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.704178|0.690618|0.677570|0.527438|0.690974|0.690960|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.704300|0.690684|0.677584|0.527515|0.691043|0.691029|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.586986|0.373828|0.274241|0.229882|0.469605|0.469596|
</details>
@@ -122,62 +117,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|12823203|155850|266302|295576135166|155850 / 295576291016|266302 / 13089505|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|8766889|3785790|4322616|295572505226|3785790 / 295576291016|4322616 / 13089505|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|8766656|3782908|4322849|295572508108|3782908 / 295576291016|4322849 / 13089505|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|3081067|321183|10008438|295575969833|321183 / 295576291016|10008438 / 13089505|
|1|Radixor|PRIMARY_OUTPUT|12781761|0|255845|288279885172|0 / 288279885172|255845 / 13037606|
|2|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|8724001|3502741|4313605|288276382431|3502741 / 288279885172|4313605 / 13037606|
|3|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|8723768|3499880|4313838|288276385292|3499880 / 288279885172|4313838 / 13037606|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|3036212|170067|10001394|288279715105|170067 / 288279885172|10001394 / 13037606|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 295576291016 (0.000000%)|13 / 13089505 (0.000099%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000100%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0.999999|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|0.999999|0.999999|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|13089492|0|13|295576291016|0 / 295576291016|13 / 13089505|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 288279885172|13 / 13037606|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|434710 / 295576291016 (0.000147%)|13 / 13089505 (0.000099%)|0.974119|0.983665|0.983796|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000100%|
</div>
@@ -185,7 +158,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.967857|0.999999|0.999999|0.999999|0.999999|0.000001|
|1|Radixor|ALL_CANDIDATES|1.000000|0.999999|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -193,15 +166,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.974119|0.983665|0.993401|0.967856|0.983797|0.983796|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|0.999999|0.999999|1.000000|1.000000|
</details>
@@ -209,7 +174,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|13089492|434710|13|295575856306|434710 / 295576291016|13 / 13089505|
|1|Radixor|ALL_CANDIDATES|13037593|0|13|288279885172|0 / 288279885172|13 / 13037606|
</details>
@@ -219,22 +184,22 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|266289|155850|278860|19162|2.492190%|4|788492|
|Radixor|255832|0|0|9613|1.265979%|4|769106|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989852|155850 / 295000681652 (0.000053%)|265613 / 13087126 (2.029575%)|0.986322|0.983830|0.983838|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.834854|3782908 / 295000681652 (0.001282%)|4322407 / 13087126 (33.027931%)|0.692561|0.683815|0.683953|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.834854|3782908 / 295000681652 (0.001282%)|4322407 / 13087126 (33.027931%)|0.692561|0.683815|0.683953|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.617630|318921 / 295000681652 (0.000108%)|10008238 / 13087126 (76.473918%)|0.577038|0.373540|0.461703|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.990213|0.000000%|1.957434%|
|2|SNOWBALL RUSSIAN DIRECT|0.834542|0.001216%|33.090302%|
|3|SNOWBALL RUSSIAN LUCENE FILTER|0.834542|0.001216%|33.090302%|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|0.616378|0.000058%|76.724356%|
</div>
@@ -242,10 +207,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987991|0.979704|0.999999|0.989852|0.999999|0.000001|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.698516|0.669721|0.999987|0.834854|0.999973|0.000027|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.698516|0.669721|0.999987|0.834854|0.999973|0.000027|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.906139|0.235261|0.999999|0.617630|0.999965|0.000035|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.980426|1.000000|0.990213|0.999999|0.000001|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.713634|0.669097|0.999988|0.834542|0.999973|0.000027|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.713634|0.669097|0.999988|0.834542|0.999973|0.000027|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.947585|0.232756|0.999999|0.616378|0.999965|0.000035|
</details>
@@ -253,21 +218,10 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986322|0.983830|0.981350|0.968175|0.983839|0.983838|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.692561|0.683815|0.675288|0.519544|0.683967|0.683953|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.692561|0.683815|0.675288|0.519544|0.683967|0.683953|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.577038|0.373540|0.276152|0.229664|0.461713|0.461703|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.983829|0.997697|0.997321|0.997509|0.997509|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.683802|0.974149|0.953634|0.963782|0.963782|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.683802|0.974149|0.953634|0.963782|0.963782|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.373528|0.994350|0.870767|0.928464|0.928464|
|1|Radixor|PRIMARY_OUTPUT|0.996023|0.990116|0.984279|0.980426|0.990164|0.990164|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|0.704259|0.690648|0.677554|0.527474|0.691007|0.690993|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|0.704259|0.690648|0.677554|0.527474|0.691007|0.690993|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|0.587020|0.373716|0.274113|0.229798|0.469634|0.469625|
</details>
@@ -275,62 +229,40 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|12821513|155850|265613|295000525802|155850 / 295000681652|265613 / 13087126|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|8764719|3782908|4322407|294996898744|3782908 / 295000681652|4322407 / 13087126|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|8764719|3782908|4322407|294996898744|3782908 / 295000681652|4322407 / 13087126|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|3078888|318921|10008238|295000362731|318921 / 295000681652|10008238 / 13087126|
|1|Radixor|PRIMARY_OUTPUT|12780071|0|255156|287711428009|0 / 287711428009|255156 / 13035227|
|2|SNOWBALL RUSSIAN DIRECT|PRIMARY_OUTPUT|8721831|3499880|4313396|287707928129|3499880 / 287711428009|4313396 / 13035227|
|3|SNOWBALL RUSSIAN LUCENE FILTER|PRIMARY_OUTPUT|8721831|3499880|4313396|287707928129|3499880 / 287711428009|4313396 / 13035227|
|4|RUSSIAN LUCENE RUSSIAN LIGHT STEM FILTER|PRIMARY_OUTPUT|3034033|167825|10001194|287711260184|167825 / 287711428009|10001194 / 13035227|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 295000681652 (0.000000%)|0 / 13087126 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|13087126|0|0|295000681652|0 / 295000681652|0 / 13087126|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 287711428009|0 / 13035227|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999999|434710 / 295000681652 (0.000147%)|0 / 13087126 (0.000000%)|0.974115|0.983663|0.983794|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -338,7 +270,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.967851|1.000000|0.999999|0.999999|0.999999|0.000001|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -346,15 +278,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.974115|0.983663|0.993401|0.967851|0.983794|0.983794|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -362,7 +286,7 @@ This mode contains **6 result rows**, **4 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|13087126|434710|0|295000246942|434710 / 295000681652|0 / 13087126|
|1|Radixor|ALL_CANDIDATES|13035227|0|0|287711428009|0 / 287711428009|0 / 13035227|
</details>
@@ -372,19 +296,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|265613|155850|278860|18991|2.472358%|4|787549|
|Radixor|255156|0|0|9442|1.244687%|4|768163|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -393,16 +317,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `RU_RU`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

395
docs/benchmarks/languages/spanish.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `ES_ES` | 65,059 | 926,393 | 120,121 | 806,272 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `es-es-default` | `1.0.0` | `ES_ES` | 65,059 | 926,393 | 120,121 | 806,272 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **926,393**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **926,393**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 5,367 | 0.579% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 524,682 | 56.637% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 240,872 | 26.001% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 130,089 | 14.043% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 25,383 | 2.740% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 8,534 | 0.921% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 522,685 | 56.422% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 243,410 | 26.275% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 124,386 | 13.427% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 27,378 | 2.955% |
## Accuracy
@@ -38,19 +38,25 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene SnowballFilter | 4.889% | 4.287% | 8.932% | Lucene TokenFilter integration path around the Snowball algorithm. |
| Official Snowball direct | 4.889% | 4.287% | 8.930% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `spanishRadixor` | 78.919 | 7.253 | 97.9 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 2079.041 | 193.548 | 2578.6 | 26.344 | Benchmark-only Spanish Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene SpanishMinimalStemFilter | `spanishLuceneSpanishMinimalStemFilter` | 45.596 | 4.639 | 56.6 | 0.578 | Minimal Spanish suffix reducer; narrow baseline. |
| Lucene SpanishLightStemFilter | `spanishLuceneSpanishLightStemFilter` | 42.003 | 1.683 | 52.1 | 0.532 | Light Spanish suffix stemmer. |
| Lucene SpanishPluralStemFilter | `spanishLuceneSpanishPluralStemFilter` | 93.734 | 6.247 | 116.3 | 1.188 | Plural-oriented Spanish suffix reducer. |
| Official Snowball direct | `snowballDirect[SPANISH]` | 171.995 | 11.035 | 213.3 | 2.179 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[SPANISH]` | 211.138 | 17.940 | 261.9 | 2.675 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `spanishRadixor` | 81.605 | 1.347 | 101.2 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 2033.430 | 14.863 | 2522.0 | 24.918 | Benchmark-only Spanish Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Lucene SpanishMinimalStemFilter | `spanishLuceneSpanishMinimalStemFilter` | 42.144 | 1.556 | 52.3 | 0.516 | Minimal Spanish suffix reducer; narrow baseline. |
| Lucene SpanishLightStemFilter | `spanishLuceneSpanishLightStemFilter` | 44.479 | 1.291 | 55.2 | 0.545 | Light Spanish suffix stemmer. |
| Lucene SpanishPluralStemFilter | `spanishLuceneSpanishPluralStemFilter` | 96.537 | 3.418 | 119.7 | 1.183 | Plural-oriented Spanish suffix reducer. |
| Official Snowball direct | `snowballDirect[SPANISH]` | 172.151 | 7.261 | 213.5 | 2.110 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[SPANISH]` | 201.697 | 9.363 | 250.2 | 2.472 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -64,33 +70,33 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `ES_ES` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `ES_ES` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/es_es/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `es-es-default`, loaded from classpath resource `org/egothor/stemmer/models/es-es-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.989295** among 7 deterministic stemmers. The runner-up is `SNOWBALL SPANISH LUCENE FILTER` at 0.652614, a difference of 0.336680. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.989429** among 7 deterministic stemmers. The runner-up is `SNOWBALL SPANISH DIRECT` at 0.652720, a difference of 0.336709. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.989448** among 7 deterministic stemmers. The runner-up is `SNOWBALL SPANISH LUCENE FILTER` at 0.652438, a difference of 0.337010. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.989580** among 7 deterministic stemmers. The runner-up is `SNOWBALL SPANISH DIRECT` at 0.652542, a difference of 0.337038. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989295|288483 / 379567318110 (0.000076%)|898652 / 41973336 (2.141007%)|0.990105|0.985755|0.985780|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.652614|2230481 / 379567318110 (0.000588%)|29161643 / 41973336 (69.476591%)|0.627151|0.449411|0.509848|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.652614|2228819 / 379567318110 (0.000587%)|29161649 / 41973336 (69.476605%)|0.627192|0.449424|0.509876|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.615102|536192 / 379567318110 (0.000141%)|32310860 / 41973336 (76.979490%)|0.583708|0.370408|0.466992|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.514823|147956 / 379567318110 (0.000039%)|40729019 / 41973336 (97.035458%)|0.130864|0.057387|0.162762|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.503874|58578 / 379567318110 (0.000015%)|41648091 / 41973336 (99.225115%)|0.037377|0.015357|0.081026|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.501768|47859 / 379567318110 (0.000013%)|41824873 / 41973336 (99.646292%)|0.017361|0.007041|0.051714|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.989448|0.000000%|2.110334%|
|2|SNOWBALL SPANISH LUCENE FILTER|0.652438|0.000414%|69.511918%|
|3|SNOWBALL SPANISH DIRECT|0.652438|0.000413%|69.511932%|
|4|HUNSPELL SPANISH LUCENE FILTER|0.615028|0.000068%|76.994273%|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|0.514565|0.000009%|97.087060%|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|0.503764|0.000002%|99.247265%|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|0.501678|0.000001%|99.664470%|
</div>
@@ -98,13 +104,13 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.993026|0.978590|0.999999|0.989295|0.999997|0.000003|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.851718|0.305234|0.999994|0.652614|0.999917|0.000083|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.851812|0.305234|0.999994|0.652614|0.999917|0.000083|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.947425|0.230205|0.999999|0.615102|0.999913|0.000087|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.893731|0.029645|1.000000|0.514823|0.999892|0.000108|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.847383|0.007749|1.000000|0.503874|0.999890|0.000110|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.756222|0.003537|1.000000|0.501768|0.999890|0.000110|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.978897|1.000000|0.989448|0.999998|0.000002|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.895438|0.304881|0.999996|0.652438|0.999915|0.000085|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.895510|0.304881|0.999996|0.652438|0.999915|0.000085|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.975281|0.230057|0.999999|0.615028|0.999910|0.000090|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.974423|0.029129|1.000000|0.514565|0.999887|0.000113|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.979154|0.007527|1.000000|0.503764|0.999885|0.000115|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.970596|0.003355|1.000000|0.501678|0.999884|0.000116|
</details>
@@ -112,27 +118,13 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.990105|0.985755|0.981443|0.971910|0.985781|0.985780|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.627151|0.449411|0.350170|0.289832|0.509876|0.509848|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.627192|0.449424|0.350173|0.289843|0.509904|0.509876|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.583708|0.370408|0.271278|0.227301|0.467014|0.466992|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.130864|0.057387|0.036752|0.029541|0.162773|0.162762|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.037377|0.015357|0.009664|0.007738|0.081032|0.081026|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.017361|0.007041|0.004416|0.003533|0.051719|0.051714|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.985753|0.995418|0.993266|0.994341|0.994341|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.449379|0.981386|0.852461|0.912391|0.912391|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.449392|0.981406|0.852463|0.912401|0.912401|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.370381|0.993314|0.790558|0.880414|0.880414|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.057381|0.993824|0.756690|0.859195|0.859195|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.015355|0.995442|0.723731|0.838115|0.838115|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.007040|0.995635|0.710610|0.829316|0.829316|
|1|Radixor|PRIMARY_OUTPUT|0.995707|0.989336|0.983046|0.978897|0.989392|0.989391|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.645406|0.454882|0.351206|0.294400|0.522496|0.522469|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.645436|0.454891|0.351208|0.294407|0.522517|0.522490|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.591847|0.372295|0.271557|0.228724|0.473678|0.473655|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.130091|0.056568|0.036142|0.029107|0.168477|0.168467|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.036514|0.014940|0.009391|0.007526|0.085851|0.085846|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.016548|0.006687|0.004191|0.003355|0.057067|0.057063|
</details>
@@ -140,71 +132,46 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|41074684|288483|898652|379567029627|288483 / 379567318110|898652 / 41973336|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|12811693|2230481|29161643|379565087629|2230481 / 379567318110|29161643 / 41973336|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|12811687|2228819|29161649|379565089291|2228819 / 379567318110|29161649 / 41973336|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|9662476|536192|32310860|379566781918|536192 / 379567318110|32310860 / 41973336|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|1244317|147956|40729019|379567170154|147956 / 379567318110|40729019 / 41973336|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|325245|58578|41648091|379567259532|58578 / 379567318110|41648091 / 41973336|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|148463|47859|41824873|379567270251|47859 / 379567318110|41824873 / 41973336|
|1|Radixor|PRIMARY_OUTPUT|41053986|0|885054|360919543590|0 / 360919543590|885054 / 41939040|
|2|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|12786409|1493087|29152631|360918050503|1493087 / 360919543590|29152631 / 41939040|
|3|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|12786403|1491944|29152637|360918051646|1491944 / 360919543590|29152637 / 41939040|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|9648381|244539|32290659|360919299051|244539 / 360919543590|32290659 / 41939040|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|1221659|32066|40717381|360919511524|32066 / 360919543590|40717381 / 41939040|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|315690|6721|41623350|360919536869|6721 / 360919543590|41623350 / 41939040|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|140718|4263|41798322|360919539327|4263 / 360919543590|41798322 / 41939040|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999993|2 / 379567318110 (0.000000%)|626 / 41973336 (0.001491%)|0.999997|0.999993|0.999993|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|0.620065|416345 / 379567318110 (0.000110%)|31894218 / 41973336 (75.986855%)|0.600268|0.384195|0.480192|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL SPANISH LUCENE FILTER|0.000062%|76.009935%|
|Radixor|0.000000%|0.001493%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0.999985|1.000000|0.999993|1.000000|0.000000|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|0.960331|0.240131|0.999999|0.620065|0.999915|0.000085|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|0.999997|0.999993|0.999988|0.999985|0.999993|0.999993|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|0.600268|0.384195|0.282504|0.237773|0.480214|0.480192|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|41972710|2|626|379567318108|2 / 379567318110|626 / 41973336|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|10079118|416345|31894218|379566901765|416345 / 379567318110|31894218 / 41973336|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL SPANISH LUCENE FILTER|223500 / 360919543590|31877837 / 41939040|
|Radixor|0 / 360919543590|626 / 41939040|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999991|1349800 / 379567318110 (0.000356%)|626 / 41973336 (0.001491%)|0.974915|0.984168|0.984289|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.620065|888077 / 379567318110 (0.000234%)|31894218 / 41973336 (75.986855%)|0.587073|0.380771|0.469749|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.999993|&lt;0.000001%|0.001493%|
|2|HUNSPELL SPANISH LUCENE FILTER|0.619950|0.000073%|76.009935%|
</div>
@@ -212,8 +179,8 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.968843|0.999985|0.999996|0.999991|0.999996|0.000004|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.919024|0.240131|0.999998|0.620065|0.999914|0.000086|
|1|Radixor|ALL_CANDIDATES|0.999959|0.999985|1.000000|0.999993|1.000000|0.000000|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.974467|0.239901|0.999999|0.619950|0.999911|0.000089|
</details>
@@ -221,17 +188,8 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.974915|0.984168|0.993598|0.968829|0.984291|0.984289|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.587073|0.380771|0.281759|0.235156|0.469773|0.469749|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999964|0.999972|0.999980|0.999944|0.999972|0.999972|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.604361|0.385016|0.282490|0.238402|0.483503|0.483480|
</details>
@@ -239,8 +197,8 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|41972710|1349800|626|379565968310|1349800 / 379567318110|626 / 41973336|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|10079118|888077|31894218|379566430033|888077 / 379567318110|31894218 / 41973336|
|1|Radixor|ALL_CANDIDATES|41938414|1737|626|360919541853|1737 / 360919543590|626 / 41939040|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|10061203|263629|31877837|360919279961|263629 / 360919543590|31877837 / 41939040|
</details>
@@ -250,26 +208,26 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL SPANISH LUCENE FILTER|416642|119847|351885|17877|2.051686%|5|890999|
|Radixor|898026|288481|1061317|42637|4.893313%|21|916797|
|HUNSPELL SPANISH LUCENE FILTER|412822|21039|19090|11309|1.331001%|5|861853|
|Radixor|884428|0|1737|20967|2.467690%|21|871404|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.989429|276044 / 377860669765 (0.000073%)|885033 / 41863370 (2.114099%)|0.990385|0.986031|0.986056|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.652720|2201196 / 377860669765 (0.000583%)|29076352 / 41863370 (69.455354%)|0.627946|0.449839|0.510450|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.652720|2201196 / 377860669765 (0.000583%)|29076352 / 41863370 (69.455354%)|0.627946|0.449839|0.510450|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.614999|531181 / 377860669765 (0.000141%)|32234855 / 41863370 (77.000144%)|0.583531|0.370163|0.466854|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.514832|146613 / 377860669765 (0.000039%)|40621522 / 41863370 (97.033569%)|0.130949|0.057424|0.162875|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.503877|57716 / 377860669765 (0.000015%)|41538714 / 41863370 (99.224487%)|0.037409|0.015370|0.081139|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.501770|47148 / 377860669765 (0.000012%)|41715144 / 41863370 (99.645929%)|0.017379|0.007049|0.051824|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.989580|&lt;0.000001%|2.084022%|
|2|SNOWBALL SPANISH DIRECT|0.652542|0.000410%|69.491126%|
|3|SNOWBALL SPANISH LUCENE FILTER|0.652542|0.000410%|69.491126%|
|4|HUNSPELL SPANISH LUCENE FILTER|0.614924|0.000068%|77.015229%|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|0.514575|0.000009%|97.085003%|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|0.503767|0.000002%|99.246590%|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|0.501679|0.000001%|99.664108%|
</div>
@@ -277,13 +235,13 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.993309|0.978859|0.999999|0.989429|0.999997|0.000003|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.853138|0.305446|0.999994|0.652720|0.999917|0.000083|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.853138|0.305446|0.999994|0.652720|0.999917|0.000083|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.947717|0.229999|0.999999|0.614999|0.999913|0.000087|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.894406|0.029664|1.000000|0.514832|0.999892|0.000108|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.849058|0.007755|1.000000|0.503877|0.999890|0.000110|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.758678|0.003541|1.000000|0.501770|0.999889|0.000111|
|1|Radixor|PRIMARY_OUTPUT|0.999999|0.979160|1.000000|0.989580|0.999998|0.000002|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.896551|0.305089|0.999996|0.652542|0.999915|0.000085|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.896551|0.305089|0.999996|0.652542|0.999915|0.000085|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.975224|0.229848|0.999999|0.614924|0.999910|0.000090|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.974539|0.029150|1.000000|0.514575|0.999887|0.000113|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.979521|0.007534|1.000000|0.503767|0.999884|0.000116|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.971230|0.003359|1.000000|0.501679|0.999884|0.000116|
</details>
@@ -291,27 +249,13 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.990385|0.986031|0.981715|0.972447|0.986057|0.986056|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.627946|0.449839|0.350441|0.290188|0.510478|0.510450|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.627946|0.449839|0.350441|0.290188|0.510478|0.510450|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.583531|0.370163|0.271053|0.227117|0.466876|0.466854|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.130949|0.057424|0.036775|0.029561|0.162886|0.162875|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.037409|0.015370|0.009672|0.007744|0.081145|0.081139|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.017379|0.007049|0.004421|0.003537|0.051829|0.051824|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986029|0.995464|0.993323|0.994392|0.994392|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.449806|0.981469|0.852556|0.912482|0.912482|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.449806|0.981469|0.852556|0.912482|0.912482|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.370136|0.993362|0.790500|0.880396|0.880396|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.057417|0.993866|0.756725|0.859234|0.859234|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.015368|0.995484|0.723753|0.838145|0.838145|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.007048|0.995676|0.710626|0.829341|0.829341|
|1|Radixor|PRIMARY_OUTPUT|0.995761|0.989470|0.983258|0.979159|0.989525|0.989523|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|0.646055|0.455257|0.351461|0.294714|0.522999|0.522972|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.646055|0.455257|0.351461|0.294714|0.522999|0.522972|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|0.591553|0.372016|0.271323|0.228513|0.473448|0.473426|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.130175|0.056607|0.036167|0.029128|0.168546|0.168536|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|0.036546|0.014953|0.009400|0.007533|0.085906|0.085901|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.016565|0.006695|0.004195|0.003359|0.057116|0.057113|
</details>
@@ -319,71 +263,46 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|40978337|276044|885033|377860393721|276044 / 377860669765|885033 / 41863370|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|12787018|2201196|29076352|377858468569|2201196 / 377860669765|29076352 / 41863370|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|12787018|2201196|29076352|377858468569|2201196 / 377860669765|29076352 / 41863370|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|9628515|531181|32234855|377860138584|531181 / 377860669765|32234855 / 41863370|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|1241848|146613|40621522|377860523152|146613 / 377860669765|40621522 / 41863370|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|324656|57716|41538714|377860612049|57716 / 377860669765|41538714 / 41863370|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|148226|47148|41715144|377860622617|47148 / 377860669765|41715144 / 41863370|
|1|Radixor|PRIMARY_OUTPUT|40958710|34|871756|359407144881|34 / 359407144915|871756 / 41830466|
|2|SNOWBALL SPANISH DIRECT|PRIMARY_OUTPUT|12762004|1472547|29068462|359405672368|1472547 / 359407144915|29068462 / 41830466|
|3|SNOWBALL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|12762004|1472547|29068462|359405672368|1472547 / 359407144915|29068462 / 41830466|
|4|HUNSPELL SPANISH LUCENE FILTER|PRIMARY_OUTPUT|9614637|244260|32215829|359406900655|244260 / 359407144915|32215829 / 41830466|
|5|SPANISH LUCENE SPANISH LIGHT STEM FILTER|PRIMARY_OUTPUT|1219357|31857|40611109|359407113058|31857 / 359407144915|40611109 / 41830466|
|6|SPANISH LUCENE SPANISH PLURAL STEM FILTER|PRIMARY_OUTPUT|315155|6589|41515311|359407138326|6589 / 359407144915|41515311 / 41830466|
|7|SPANISH LUCENE SPANISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|140505|4162|41689961|359407140753|4162 / 359407144915|41689961 / 41830466|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 377860669765 (0.000000%)|0 / 41863370 (0.000000%)|1.000000|1.000000|1.000000|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|0.619928|412198 / 377860669765 (0.000109%)|31822108 / 41863370 (76.014205%)|0.600000|0.383864|0.479978|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL SPANISH LUCENE FILTER|0.000062%|76.037484%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|0.960568|0.239858|0.999999|0.619928|0.999915|0.000085|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|0.600000|0.383864|0.282205|0.237519|0.480000|0.479978|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|41863370|0|0|377860669765|0 / 377860669765|0 / 41863370|
|2|HUNSPELL SPANISH LUCENE FILTER|ANY_CANDIDATE|10041262|412198|31822108|377860257567|412198 / 377860669765|31822108 / 41863370|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL SPANISH LUCENE FILTER|223274 / 359407144915|31806834 / 41830466|
|Radixor|0 / 359407144915|0 / 41830466|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999998|1255381 / 377860669765 (0.000332%)|0 / 41863370 (0.000000%)|0.976572|0.985228|0.985334|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.619928|878949 / 377860669765 (0.000233%)|31822108 / 41863370 (76.014205%)|0.586905|0.380469|0.469606|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|&lt;0.000001%|0.000000%|
|2|HUNSPELL SPANISH LUCENE FILTER|0.619812|0.000073%|76.037484%|
</div>
@@ -391,8 +310,8 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.970885|1.000000|0.999997|0.999998|0.999997|0.000003|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.919512|0.239858|0.999998|0.619928|0.999913|0.000087|
|1|Radixor|ALL_CANDIDATES|0.999987|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.974405|0.239625|0.999999|0.619812|0.999911|0.000089|
</details>
@@ -400,17 +319,8 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.976572|0.985228|0.994038|0.970885|0.985335|0.985334|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.586905|0.380469|0.281467|0.234926|0.469630|0.469606|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|0.999989|0.999993|0.999997|0.999987|0.999993|0.999993|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|0.603992|0.384656|0.282183|0.238126|0.483210|0.483187|
</details>
@@ -418,8 +328,8 @@ This mode contains **11 result rows**, **7 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|41863370|1255381|0|377859414384|1255381 / 377860669765|0 / 41863370|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|10041262|878949|31822108|377859790816|878949 / 377860669765|31822108 / 41863370|
|1|Radixor|ALL_CANDIDATES|41830466|560|0|359407144355|560 / 359407144915|0 / 41830466|
|2|HUNSPELL SPANISH LUCENE FILTER|ALL_CANDIDATES|10023632|263289|31806834|359406881626|263289 / 359407144915|31806834 / 41830466|
</details>
@@ -429,20 +339,20 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL SPANISH LUCENE FILTER|412747|118983|347768|17807|2.048262%|5|888962|
|Radixor|885033|276044|979337|42403|4.877434%|21|914127|
|HUNSPELL SPANISH LUCENE FILTER|408995|20986|19029|11287|1.331204%|5|860048|
|Radixor|871756|34|526|20911|2.466272%|21|869542|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -451,16 +361,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `ES_ES`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

319
docs/benchmarks/languages/swedish.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `SV_SE` | 12,371 | 110,468 | 24,731 | 85,737 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `sv-se-default` | `1.0.0` | `SV_SE` | 12,371 | 110,468 | 24,731 | 85,737 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **110,468**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **110,468**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 502 | 0.454% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 14,268 | 12.916% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 66,796 | 60.466% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 25,745 | 23.305% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 3,157 | 2.858% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 711 | 0.644% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 14,126 | 12.787% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 68,749 | 62.234% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 23,583 | 21.348% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 3,299 | 2.986% |
## Accuracy
@@ -36,17 +36,23 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Official Snowball direct | 40.068% | 37.512% | 48.926% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
| Lucene SnowballFilter | 38.785% | 35.839% | 48.999% | Lucene TokenFilter integration path around the Snowball algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `swedishRadixor` | 5.489 | 0.355 | 64.0 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene SwedishMinimalStemFilter | `swedishLuceneSwedishMinimalStemFilter` | 4.630 | 0.130 | 54.0 | 0.843 | Minimal Swedish suffix reducer. |
| Lucene SwedishLightStemFilter | `swedishLuceneSwedishLightStemFilter` | 4.876 | 0.328 | 56.9 | 0.888 | Light Swedish suffix stemmer. |
| Official Snowball direct | `snowballDirect[SWEDISH]` | 7.517 | 0.072 | 87.7 | 1.370 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[SWEDISH]` | 9.793 | 0.338 | 114.2 | 1.784 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `swedishRadixor` | 5.476 | 0.081 | 63.9 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene SwedishMinimalStemFilter | `swedishLuceneSwedishMinimalStemFilter` | 4.741 | 0.086 | 55.3 | 0.866 | Minimal Swedish suffix reducer. |
| Lucene SwedishLightStemFilter | `swedishLuceneSwedishLightStemFilter` | 4.893 | 0.053 | 57.1 | 0.893 | Light Swedish suffix stemmer. |
| Official Snowball direct | `snowballDirect[SWEDISH]` | 7.606 | 0.555 | 88.7 | 1.389 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[SWEDISH]` | 10.295 | 0.749 | 120.1 | 1.880 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -60,31 +66,31 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `SV_SE` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `SV_SE` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/sv_se/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `sv-se-default`, loaded from classpath resource `org/egothor/stemmer/models/sv-se-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.974636** among 5 deterministic stemmers. The runner-up is `SNOWBALL SWEDISH DIRECT` at 0.807534, a difference of 0.167101. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.974584** among 5 deterministic stemmers. The runner-up is `SNOWBALL SWEDISH DIRECT` at 0.807599, a difference of 0.166985. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.977619** among 5 deterministic stemmers. The runner-up is `SNOWBALL SWEDISH DIRECT` at 0.808543, a difference of 0.169076. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.977573** among 5 deterministic stemmers. The runner-up is `SNOWBALL SWEDISH DIRECT` at 0.808611, a difference of 0.168961. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.974636|24473 / 4812155436 (0.000509%)|19546 / 385342 (5.072377%)|0.939665|0.943246|0.943260|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.807534|67105 / 4812155436 (0.001394%)|148325 / 385342 (38.491781%)|0.739832|0.687540|0.692339|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.799307|64262 / 4812155436 (0.001335%)|154666 / 385342 (40.137333%)|0.736940|0.678180|0.684227|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.796072|40227 / 4812155436 (0.000836%)|157161 / 385342 (40.784809%)|0.781991|0.698068|0.709491|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.783685|45941 / 4812155436 (0.000955%)|166707 / 385342 (43.262089%)|0.757232|0.672808|0.684713|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.977619|0.000000%|4.476263%|
|2|SNOWBALL SWEDISH DIRECT|0.808543|0.000821%|38.290570%|
|3|SNOWBALL SWEDISH LUCENE FILTER|0.800222|0.000775%|39.954747%|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|0.797907|0.000439%|40.418073%|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|0.785227|0.000534%|42.954113%|
</div>
@@ -92,11 +98,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.937292|0.949276|0.999995|0.974636|0.999991|0.000009|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.779348|0.615082|0.999986|0.807534|0.999955|0.000045|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.782117|0.598627|0.999987|0.799307|0.999955|0.000045|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.850127|0.592152|0.999992|0.796072|0.999959|0.000041|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.826360|0.567379|0.999990|0.783685|0.999956|0.000044|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.955237|1.000000|0.977619|0.999996|0.000004|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.863080|0.617094|0.999992|0.808543|0.999960|0.000040|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.866630|0.600453|0.999992|0.800222|0.999959|0.000041|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.919176|0.595819|0.999996|0.797907|0.999962|0.000038|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.899588|0.570459|0.999995|0.785227|0.999959|0.000041|
</details>
@@ -104,23 +110,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.939665|0.943246|0.946855|0.892588|0.943265|0.943260|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.739832|0.687540|0.642152|0.523856|0.692361|0.692339|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.736940|0.678180|0.628098|0.513065|0.684249|0.684227|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.781991|0.698068|0.630412|0.536179|0.709510|0.709491|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.757232|0.672808|0.605321|0.506941|0.684733|0.684713|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.943241|0.992631|0.993395|0.993013|0.993013|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.687518|0.984860|0.942685|0.963311|0.963311|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.678157|0.985207|0.939659|0.961894|0.961894|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.698048|0.988493|0.944582|0.966038|0.966038|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.672787|0.986795|0.942303|0.964036|0.964036|
|1|Radixor|PRIMARY_OUTPUT|0.990715|0.977106|0.963866|0.955237|0.977362|0.977361|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.799353|0.719647|0.654396|0.562070|0.729796|0.729777|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.796053|0.709394|0.639751|0.549660|0.721367|0.721348|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.829176|0.722989|0.640913|0.566158|0.740043|0.740026|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.806522|0.698179|0.615497|0.536309|0.716364|0.716347|
</details>
@@ -128,63 +122,41 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|365796|24473|19546|4812130963|24473 / 4812155436|19546 / 385342|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|237017|67105|148325|4812088331|67105 / 4812155436|148325 / 385342|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|230676|64262|154666|4812091174|64262 / 4812155436|154666 / 385342|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|228181|40227|157161|4812115209|40227 / 4812155436|157161 / 385342|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|218635|45941|166707|4812109495|45941 / 4812155436|166707 / 385342|
|1|Radixor|PRIMARY_OUTPUT|362653|0|16994|4529284143|0 / 4529284143|16994 / 379647|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|234278|37166|145369|4529246977|37166 / 4529284143|145369 / 379647|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|227960|35082|151687|4529249061|35082 / 4529284143|151687 / 379647|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|226201|19890|153446|4529264253|19890 / 4529284143|153446 / 379647|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|216573|24174|163074|4529259969|24174 / 4529284143|163074 / 379647|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 4812155436 (0.000000%)|0 / 385342 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|385342|0|0|4812155436|0 / 4812155436|0 / 385342|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 4529284143|0 / 379647|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999995|47848 / 4812155436 (0.000994%)|0 / 385342 (0.000000%)|0.909640|0.941544|0.943152|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -192,7 +164,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.889545|1.000000|0.999990|0.999995|0.999990|0.000010|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -200,15 +172,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.909640|0.941544|0.975768|0.889545|0.943157|0.943152|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -216,7 +180,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|385342|47848|0|4812107588|47848 / 4812155436|0 / 385342|
|1|Radixor|ALL_CANDIDATES|379647|0|0|4529284143|0 / 4529284143|0 / 379647|
</details>
@@ -226,23 +190,23 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|19546|24473|23375|5767|5.878216%|5|104148|
|Radixor|16994|0|0|2840|2.983789%|5|98108|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.974584|24473 / 4789911577 (0.000511%)|19546 / 384563 (5.082652%)|0.939544|0.943132|0.943146|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.807599|67105 / 4789911577 (0.001401%)|147975 / 384563 (38.478741%)|0.739645|0.687500|0.692274|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.799355|64262 / 4789911577 (0.001342%)|154316 / 384563 (40.127625%)|0.736744|0.678122|0.684143|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.795947|40227 / 4789911577 (0.000840%)|156939 / 384563 (40.809698%)|0.781694|0.697790|0.709212|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.783598|45941 / 4789911577 (0.000959%)|166437 / 384563 (43.279515%)|0.756945|0.672575|0.684469|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.977573|0.000000%|4.485467%|
|2|SNOWBALL SWEDISH DIRECT|0.808611|0.000824%|38.276920%|
|3|SNOWBALL SWEDISH LUCENE FILTER|0.800274|0.000778%|39.944519%|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|0.797785|0.000441%|40.442582%|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|0.785141|0.000536%|42.971167%|
</div>
@@ -250,11 +214,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.937167|0.949173|0.999995|0.974584|0.999991|0.000009|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.779037|0.615213|0.999986|0.807599|0.999955|0.000045|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.781800|0.598724|0.999987|0.799355|0.999954|0.000046|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.849816|0.591903|0.999992|0.795947|0.999959|0.000041|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.826025|0.567205|0.999990|0.783598|0.999956|0.000044|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.955145|1.000000|0.977573|0.999996|0.000004|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.862864|0.617231|0.999992|0.808611|0.999960|0.000040|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.866412|0.600555|0.999992|0.800274|0.999959|0.000041|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.918993|0.595574|0.999996|0.797785|0.999962|0.000038|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.899375|0.570288|0.999995|0.785141|0.999959|0.000041|
</details>
@@ -262,23 +226,11 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.939544|0.943132|0.946748|0.892384|0.943151|0.943146|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.739645|0.687500|0.642223|0.523810|0.692296|0.692274|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.736744|0.678122|0.628142|0.512999|0.684165|0.684143|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.781694|0.697790|0.630152|0.535851|0.709231|0.709212|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.756945|0.672575|0.605126|0.506676|0.684489|0.684469|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.943127|0.992612|0.993378|0.992995|0.992995|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.687478|0.984821|0.942695|0.963298|0.963298|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.678100|0.985169|0.939661|0.961877|0.961877|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.697770|0.988463|0.944528|0.965996|0.965996|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.672553|0.986761|0.942265|0.964000|0.964000|
|1|Radixor|PRIMARY_OUTPUT|0.990695|0.977058|0.963791|0.955145|0.977315|0.977314|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|0.799250|0.719665|0.654494|0.562091|0.729785|0.729766|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|0.795941|0.709393|0.639820|0.549658|0.721337|0.721319|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|0.828962|0.722752|0.640668|0.565867|0.739816|0.739800|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|0.806317|0.697987|0.615318|0.536083|0.716172|0.716155|
</details>
@@ -286,63 +238,41 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|365017|24473|19546|4789887104|24473 / 4789911577|19546 / 384563|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|236588|67105|147975|4789844472|67105 / 4789911577|147975 / 384563|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|230247|64262|154316|4789847315|64262 / 4789911577|154316 / 384563|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|227624|40227|156939|4789871350|40227 / 4789911577|156939 / 384563|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|218126|45941|166437|4789865636|45941 / 4789911577|166437 / 384563|
|1|Radixor|PRIMARY_OUTPUT|361874|0|16994|4507704713|0 / 4507704713|16994 / 378868|
|2|SNOWBALL SWEDISH DIRECT|PRIMARY_OUTPUT|233849|37166|145019|4507667547|37166 / 4507704713|145019 / 378868|
|3|SNOWBALL SWEDISH LUCENE FILTER|PRIMARY_OUTPUT|227531|35082|151337|4507669631|35082 / 4507704713|151337 / 378868|
|4|SWEDISH LUCENE SWEDISH MINIMAL STEM FILTER|PRIMARY_OUTPUT|225644|19890|153224|4507684823|19890 / 4507704713|153224 / 378868|
|5|SWEDISH LUCENE SWEDISH LIGHT STEM FILTER|PRIMARY_OUTPUT|216064|24174|162804|4507680539|24174 / 4507704713|162804 / 378868|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 4789911577 (0.000000%)|0 / 384563 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|384563|0|0|4789911577|0 / 4789911577|0 / 384563|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 4507704713|0 / 378868|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999995|47848 / 4789911577 (0.000999%)|0 / 384563 (0.000000%)|0.909473|0.941433|0.943047|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -350,7 +280,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.889346|1.000000|0.999990|0.999995|0.999990|0.000010|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -358,15 +288,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.909473|0.941433|0.975720|0.889346|0.943051|0.943047|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -374,7 +296,7 @@ This mode contains **7 result rows**, **5 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|384563|47848|0|4789863729|47848 / 4789911577|0 / 384563|
|1|Radixor|ALL_CANDIDATES|378868|0|0|4507704713|0 / 4507704713|0 / 378868|
</details>
@@ -384,19 +306,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|19546|24473|23375|5767|5.891848%|5|103921|
|Radixor|16994|0|0|2840|2.990922%|5|97881|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -405,16 +327,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `SV_SE`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

407
docs/benchmarks/languages/ukrainian.md
View File

@@ -8,21 +8,21 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `UK_UA` | 1,493 | 15,737 | 2,985 | 12,752 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `uk-ua-default` | `1.0.0` | `UK_UA` | 1,493 | 15,737 | 2,985 | 12,752 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **15,737**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **15,737**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 249 | 1.582% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 4,160 | 26.435% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 5,859 | 37.231% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 3,004 | 19.089% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 2,465 | 15.664% |
| `AppendCharacterCommand` | Appends one character to the end of the word form. | 267 | 1.697% |
| `BackwardCompoundCommand` | Applies a multi-step backward patch made from skip, delete, insert, and replace operations. | 4,156 | 26.409% |
| `DeleteSuffixCommand` | Deletes one or more trailing characters from the word form. | 5,883 | 37.383% |
| `PreserveCommand` | Returns the word form unchanged because it already matches the preferred root. | 2,962 | 18.822% |
| `ReplaceLastCharacterCommand` | Replaces the final character of the word form. | 2,469 | 15.689% |
## Accuracy
@@ -35,16 +35,22 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene MorfologikFilter | 92.362% | 90.637% | 99.732% | Dictionary-based path; Morfologik can emit multiple terms. |
| Morfologik direct | 92.362% | 90.637% | 99.732% | Direct dictionary lookup; first returned stem is used for quality when no ranking weight is exposed. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `ukrainianRadixor` | 0.682 | 0.057 | 53.5 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 43.527 | 1.207 | 3413.3 | 63.799 | Benchmark-only Ukrainian Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Morfologik direct | `ukrainianMorfologikDirect` | 8.680 | 0.073 | 680.7 | 12.723 | Direct Morfologik dictionary lookup; first returned stem is used for quality. |
| Lucene MorfologikFilter | `ukrainianLuceneMorfologikFilter` | 14.575 | 0.248 | 1143.0 | 21.364 | Dictionary-based Morfologik TokenFilter; may emit multiple terms. |
| Radixor | `ukrainianRadixor` | 0.639 | 0.009 | 50.1 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Lucene HunspellStemFilter | `luceneHunspellStemFilter` | 47.919 | 5.308 | 3757.8 | 74.957 | Benchmark-only Ukrainian Hunspell dictionary compared via Lucene HunspellStemFilter. |
| Morfologik direct | `ukrainianMorfologikDirect` | 8.662 | 0.105 | 679.3 | 13.550 | Direct Morfologik dictionary lookup; first returned stem is used for quality. |
| Lucene MorfologikFilter | `ukrainianLuceneMorfologikFilter` | 15.367 | 0.219 | 1205.1 | 24.038 | Dictionary-based Morfologik TokenFilter; may emit multiple terms. |
## Interpretation Notes
@@ -58,30 +64,30 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `UK_UA` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `UK_UA` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/uk_ua/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `uk-ua-default`, loaded from classpath resource `org/egothor/stemmer/models/uk-ua-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.995343** among 4 deterministic stemmers. The runner-up is `UKRAINIAN LUCENE MORFOLOGIK FILTER` at 0.928768, a difference of 0.066575. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.995342** among 4 deterministic stemmers. The runner-up is `UKRAINIAN LUCENE MORFOLOGIK FILTER` at 0.928751, a difference of 0.066591. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.995816** among 4 deterministic stemmers. The runner-up is `UKRAINIAN LUCENE MORFOLOGIK FILTER` at 0.928906, a difference of 0.066910. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.995815** among 4 deterministic stemmers. The runner-up is `UKRAINIAN LUCENE MORFOLOGIK FILTER` at 0.928888, a difference of 0.066926. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.995343|880 / 101387550 (0.000868%)|608 / 65340 (0.930517%)|0.987406|0.988637|0.988632|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.928768|828 / 101387550 (0.000817%)|9308 / 65340 (14.245485%)|0.956896|0.917054|0.919223|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.928646|828 / 101387550 (0.000817%)|9324 / 65340 (14.269972%)|0.956832|0.916912|0.919090|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.885793|794 / 101387550 (0.000783%)|14924 / 65340 (22.840526%)|0.933008|0.865139|0.871499|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.995816|0.000000%|0.836852%|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|0.928906|0.000028%|14.218810%|
|3|UKRAINIAN MORFOLOGIK DIRECT|0.928783|0.000028%|14.243378%|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|0.885789|0.000006%|22.842226%|
</div>
@@ -89,10 +95,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986588|0.990695|0.999991|0.995343|0.999985|0.000015|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.985438|0.857545|0.999992|0.928768|0.999900|0.000100|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.985434|0.857300|0.999992|0.928646|0.999900|0.000100|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.984495|0.771595|0.999992|0.885793|0.999845|0.000155|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.991631|1.000000|0.995816|0.999995|0.000005|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.999499|0.857812|1.000000|0.928906|0.999907|0.000093|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.999499|0.857566|1.000000|0.928783|0.999907|0.000093|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.999881|0.771578|1.000000|0.885789|0.999851|0.000149|
</details>
@@ -100,21 +106,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987406|0.988637|0.989871|0.977529|0.988639|0.988632|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.956896|0.917054|0.880397|0.846814|0.919270|0.919223|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.956832|0.916912|0.880190|0.846572|0.919137|0.919090|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.933008|0.865139|0.806475|0.762331|0.871568|0.871499|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988630|0.997994|0.998266|0.998130|0.998130|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.917004|0.997990|0.971000|0.984310|0.984310|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.916862|0.997990|0.970876|0.984246|0.984246|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.865063|0.998114|0.949804|0.973360|0.973360|
|1|Radixor|PRIMARY_OUTPUT|0.998315|0.995798|0.993294|0.991631|0.995807|0.995804|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.967537|0.923251|0.882842|0.857443|0.925949|0.925906|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.967474|0.923109|0.882634|0.857198|0.925817|0.925774|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.944015|0.871018|0.808499|0.771507|0.878343|0.878277|
</details>
@@ -122,80 +117,49 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|64732|880|608|101386670|880 / 101387550|608 / 65340|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|56032|828|9308|101386722|828 / 101387550|9308 / 65340|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|56016|828|9324|101386722|828 / 101387550|9324 / 65340|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|50416|794|14924|101386756|794 / 101387550|14924 / 65340|
|1|Radixor|PRIMARY_OUTPUT|64580|0|545|100039050|0 / 100039050|545 / 65125|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|55865|28|9260|100039022|28 / 100039050|9260 / 65125|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|55849|28|9276|100039022|28 / 100039050|9276 / 65125|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|50249|6|14876|100039044|6 / 100039050|14876 / 65125|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 101387550 (0.000000%)|0 / 65340 (0.000000%)|1.000000|1.000000|1.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.962151|122 / 101387550 (0.000120%)|4946 / 65340 (7.569636%)|0.982323|0.959732|0.960413|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|0.962029|122 / 101387550 (0.000120%)|4962 / 65340 (7.594123%)|0.982267|0.959599|0.960286|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|0.927570|326 / 101387550 (0.000322%)|9465 / 65340 (14.485767%)|0.962884|0.919443|0.922008|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL UKRAINIAN LUCENE FILTER|0.000000%|14.533589%|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|0.000000%|7.594626%|
|UKRAINIAN MORFOLOGIK DIRECT|0.000000%|7.619194%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.997984|0.924304|0.999999|0.962151|0.999950|0.000050|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|0.997983|0.924059|0.999999|0.962029|0.999950|0.000050|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|0.994199|0.855142|0.999997|0.927570|0.999903|0.000097|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.982323|0.959732|0.938156|0.922581|0.960438|0.960413|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|0.982267|0.959599|0.937954|0.922337|0.960310|0.960286|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|0.962884|0.919443|0.879752|0.850897|0.922053|0.922008|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|65340|0|0|101387550|0 / 101387550|0 / 65340|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|60394|122|4946|101387428|122 / 101387550|4946 / 65340|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|60378|122|4962|101387428|122 / 101387550|4962 / 65340|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|55875|326|9465|101387224|326 / 101387550|9465 / 65340|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL UKRAINIAN LUCENE FILTER|0 / 100039050|9465 / 65125|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|0 / 100039050|4946 / 65125|
|UKRAINIAN MORFOLOGIK DIRECT|0 / 100039050|4962 / 65125|
|Radixor|0 / 100039050|0 / 65125|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999993|1490 / 101387550 (0.001470%)|0 / 65340 (0.000000%)|0.982084|0.988727|0.988782|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.962145|1368 / 101387550 (0.001349%)|4946 / 65340 (7.569636%)|0.966650|0.950323|0.950669|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.962023|1368 / 101387550 (0.001349%)|4962 / 65340 (7.594123%)|0.966592|0.950191|0.950541|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.927565|1271 / 101387550 (0.001254%)|9465 / 65340 (14.485767%)|0.950501|0.912349|0.914347|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|0.962027|0.000059%|7.594626%|
|3|UKRAINIAN MORFOLOGIK DIRECT|0.961904|0.000059%|7.619194%|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|0.927332|0.000047%|14.533589%|
</div>
@@ -203,10 +167,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.977705|1.000000|0.999985|0.999993|0.999985|0.000015|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.977850|0.924304|0.999987|0.962145|0.999938|0.000062|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.977845|0.924059|0.999987|0.962023|0.999938|0.000062|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.977759|0.855142|0.999987|0.927565|0.999894|0.000106|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.999021|0.924054|0.999999|0.962027|0.999950|0.000050|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.999020|0.923808|0.999999|0.961904|0.999950|0.000050|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.999156|0.854664|1.000000|0.927332|0.999905|0.000095|
</details>
@@ -214,21 +178,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.982084|0.988727|0.995460|0.977705|0.988789|0.988782|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.966650|0.950323|0.934539|0.905349|0.950700|0.950669|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.966592|0.950191|0.934337|0.905109|0.950571|0.950541|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.950501|0.912349|0.877142|0.838825|0.914398|0.914347|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.983070|0.960076|0.938133|0.923217|0.960806|0.960782|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.983014|0.959943|0.937931|0.922972|0.960678|0.960654|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.966477|0.921279|0.880120|0.854048|0.924090|0.924046|
</details>
@@ -236,10 +189,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|65340|1490|0|101386060|1490 / 101387550|0 / 65340|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|60394|1368|4946|101386182|1368 / 101387550|4946 / 65340|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|60378|1368|4962|101386182|1368 / 101387550|4962 / 65340|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|55875|1271|9465|101386279|1271 / 101387550|9465 / 65340|
|1|Radixor|ALL_CANDIDATES|65125|0|0|100039050|0 / 100039050|0 / 65125|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|60179|59|4946|100038991|59 / 100039050|4946 / 65125|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|60163|59|4962|100038991|59 / 100039050|4962 / 65125|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|55660|47|9465|100039003|47 / 100039050|9465 / 65125|
</details>
@@ -249,25 +202,25 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL UKRAINIAN LUCENE FILTER|5459|468|477|1322|9.280449%|6|15740|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|4362|706|540|2207|15.493155%|6|16937|
|UKRAINIAN MORFOLOGIK DIRECT|4362|706|540|2207|15.493155%|6|16937|
|Radixor|608|880|610|190|1.333801%|2|14435|
|HUNSPELL UKRAINIAN LUCENE FILTER|5411|6|41|1259|8.897527%|6|15577|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|4314|28|31|2130|15.053004%|6|16748|
|UKRAINIAN MORFOLOGIK DIRECT|4314|28|31|2130|15.053004%|6|16748|
|Radixor|545|0|0|95|0.671378%|2|14245|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.995342|880 / 101259406 (0.000869%)|608 / 65324 (0.930745%)|0.987403|0.988634|0.988629|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.928751|828 / 101259406 (0.000818%)|9308 / 65324 (14.248974%)|0.956884|0.917032|0.919202|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.928751|828 / 101259406 (0.000818%)|9308 / 65324 (14.248974%)|0.956884|0.917032|0.919202|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.885796|794 / 101259406 (0.000784%)|14920 / 65324 (22.839998%)|0.933007|0.865141|0.871500|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.995815|0.000000%|0.837058%|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|0.928888|0.000028%|14.222304%|
|3|UKRAINIAN MORFOLOGIK DIRECT|0.928888|0.000028%|14.222304%|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|0.885791|0.000006%|22.841696%|
</div>
@@ -275,10 +228,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.986585|0.990693|0.999991|0.995342|0.999985|0.000015|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.985434|0.857510|0.999992|0.928751|0.999900|0.000100|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.985434|0.857510|0.999992|0.928751|0.999900|0.000100|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.984492|0.771600|0.999992|0.885796|0.999845|0.000155|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.991629|1.000000|0.995815|0.999995|0.000005|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.999499|0.857777|1.000000|0.928888|0.999907|0.000093|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.999499|0.857777|1.000000|0.928888|0.999907|0.000093|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.999881|0.771583|1.000000|0.885791|0.999851|0.000149|
</details>
@@ -286,21 +239,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.987403|0.988634|0.989868|0.977524|0.988636|0.988629|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.956884|0.917032|0.880367|0.846777|0.919249|0.919202|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.956884|0.917032|0.880367|0.846777|0.919249|0.919202|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.933007|0.865141|0.806479|0.762334|0.871570|0.871500|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988627|0.997992|0.998264|0.998128|0.998128|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.916982|0.997988|0.970978|0.984298|0.984298|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.916982|0.997988|0.970978|0.984298|0.984298|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.865065|0.998113|0.949788|0.973351|0.973351|
|1|Radixor|PRIMARY_OUTPUT|0.998315|0.995797|0.993292|0.991629|0.995806|0.995803|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|0.967528|0.923231|0.882812|0.857408|0.925930|0.925887|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|0.967528|0.923231|0.882812|0.857408|0.925930|0.925887|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|0.944017|0.871021|0.808503|0.771512|0.878346|0.878280|
</details>
@@ -308,80 +250,49 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|64716|880|608|101258526|880 / 101259406|608 / 65324|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|56016|828|9308|101258578|828 / 101259406|9308 / 65324|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|56016|828|9308|101258578|828 / 101259406|9308 / 65324|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|50404|794|14920|101258612|794 / 101259406|14920 / 65324|
|1|Radixor|PRIMARY_OUTPUT|64564|0|545|99911761|0 / 99911761|545 / 65109|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|PRIMARY_OUTPUT|55849|28|9260|99911733|28 / 99911761|9260 / 65109|
|3|UKRAINIAN MORFOLOGIK DIRECT|PRIMARY_OUTPUT|55849|28|9260|99911733|28 / 99911761|9260 / 65109|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|PRIMARY_OUTPUT|50237|6|14872|99911755|6 / 99911761|14872 / 65109|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 101259406 (0.000000%)|0 / 65324 (0.000000%)|1.000000|1.000000|1.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.962142|122 / 101259406 (0.000120%)|4946 / 65324 (7.571490%)|0.982318|0.959722|0.960404|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|0.962142|122 / 101259406 (0.000120%)|4946 / 65324 (7.571490%)|0.982318|0.959722|0.960404|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|0.927552|326 / 101259406 (0.000322%)|9465 / 65324 (14.489315%)|0.962874|0.919422|0.921988|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|HUNSPELL UKRAINIAN LUCENE FILTER|0.000000%|14.537161%|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|0.000000%|7.596492%|
|UKRAINIAN MORFOLOGIK DIRECT|0.000000%|7.596492%|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.997983|0.924285|0.999999|0.962142|0.999950|0.000050|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|0.997983|0.924285|0.999999|0.962142|0.999950|0.000050|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|0.994198|0.855107|0.999997|0.927552|0.999903|0.000097|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|0.982318|0.959722|0.938141|0.922562|0.960428|0.960404|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|0.982318|0.959722|0.938141|0.922562|0.960428|0.960404|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|0.962874|0.919422|0.879722|0.850861|0.922033|0.921988|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|65324|0|0|101259406|0 / 101259406|0 / 65324|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ANY_CANDIDATE|60378|122|4946|101259284|122 / 101259406|4946 / 65324|
|3|UKRAINIAN MORFOLOGIK DIRECT|ANY_CANDIDATE|60378|122|4946|101259284|122 / 101259406|4946 / 65324|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ANY_CANDIDATE|55859|326|9465|101259080|326 / 101259406|9465 / 65324|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|HUNSPELL UKRAINIAN LUCENE FILTER|0 / 99911761|9465 / 65109|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|0 / 99911761|4946 / 65109|
|UKRAINIAN MORFOLOGIK DIRECT|0 / 99911761|4946 / 65109|
|Radixor|0 / 99911761|0 / 65109|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999993|1490 / 101259406 (0.001471%)|0 / 65324 (0.000000%)|0.982079|0.988724|0.988779|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.962136|1368 / 101259406 (0.001351%)|4946 / 65324 (7.571490%)|0.966642|0.950311|0.950657|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.962136|1368 / 101259406 (0.001351%)|4946 / 65324 (7.571490%)|0.966642|0.950311|0.950657|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.927547|1271 / 101259406 (0.001255%)|9465 / 65324 (14.489315%)|0.950487|0.912326|0.914325|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|0.962017|0.000059%|7.596492%|
|3|UKRAINIAN MORFOLOGIK DIRECT|0.962017|0.000059%|7.596492%|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|0.927314|0.000047%|14.537161%|
</div>
@@ -389,10 +300,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.977699|1.000000|0.999985|0.999993|0.999985|0.000015|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.977845|0.924285|0.999986|0.962136|0.999938|0.000062|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.977845|0.924285|0.999986|0.962136|0.999938|0.000062|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.977752|0.855107|0.999987|0.927547|0.999894|0.000106|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.999020|0.924035|0.999999|0.962017|0.999950|0.000050|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.999020|0.924035|0.999999|0.962017|0.999950|0.000050|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.999156|0.854628|1.000000|0.927314|0.999905|0.000095|
</details>
@@ -400,21 +311,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.982079|0.988724|0.995459|0.977699|0.988787|0.988779|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.966642|0.950311|0.934522|0.905326|0.950688|0.950657|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.966642|0.950311|0.934522|0.905326|0.950688|0.950657|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.950487|0.912326|0.877111|0.838787|0.914376|0.914325|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|0.983065|0.960066|0.938118|0.923199|0.960796|0.960772|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|0.983065|0.960066|0.938118|0.923199|0.960796|0.960772|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|0.966468|0.921258|0.880089|0.854012|0.924071|0.924027|
</details>
@@ -422,10 +322,10 @@ This mode contains **12 result rows**, **4 evaluated stemmers**, and **3 output
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|65324|1490|0|101257916|1490 / 101259406|0 / 65324|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|60378|1368|4946|101258038|1368 / 101259406|4946 / 65324|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|60378|1368|4946|101258038|1368 / 101259406|4946 / 65324|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|55859|1271|9465|101258135|1271 / 101259406|9465 / 65324|
|1|Radixor|ALL_CANDIDATES|65109|0|0|99911761|0 / 99911761|0 / 65109|
|2|UKRAINIAN LUCENE MORFOLOGIK FILTER|ALL_CANDIDATES|60163|59|4946|99911702|59 / 99911761|4946 / 65109|
|3|UKRAINIAN MORFOLOGIK DIRECT|ALL_CANDIDATES|60163|59|4946|99911702|59 / 99911761|4946 / 65109|
|4|HUNSPELL UKRAINIAN LUCENE FILTER|ALL_CANDIDATES|55644|47|9465|99911714|47 / 99911761|9465 / 65109|
</details>
@@ -435,22 +335,22 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|HUNSPELL UKRAINIAN LUCENE FILTER|5455|468|477|1321|9.279292%|6|15730|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|4362|706|540|2207|15.502950%|6|16928|
|UKRAINIAN MORFOLOGIK DIRECT|4362|706|540|2207|15.502950%|6|16928|
|Radixor|608|880|610|190|1.334645%|2|14426|
|HUNSPELL UKRAINIAN LUCENE FILTER|5407|6|41|1258|8.896118%|6|15567|
|UKRAINIAN LUCENE MORFOLOGIK FILTER|4314|28|31|2130|15.062584%|6|16739|
|UKRAINIAN MORFOLOGIK DIRECT|4314|28|31|2130|15.062584%|6|16739|
|Radixor|545|0|0|95|0.671805%|2|14236|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -459,16 +359,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `UK_UA`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->

269
docs/benchmarks/languages/yiddish.md
View File

@@ -8,13 +8,13 @@ Radixor must not be read as simply "slower" when a narrow competitor has a lower
## Dictionary Corpus
| Resource | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | ---: | ---: | ---: | ---: |
| `YI` | 802 | 4,300 | 1,524 | 2,776 |
| Model ID | Model version | Language | Dictionary rows | Complete quality tokens | Already-root tokens | Changed speed tokens |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| `yi-default` | `1.0.0` | `YI` | 802 | 4,300 | 1,524 | 2,776 |
## Radixor Patch Command Distribution
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete language dictionary. The total number of preferred patch commands analyzed for this language is **4,300**.
Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is **4,300**.
| Command class | Meaning | Word forms | Share |
| --- | --- | ---: | ---: |
@@ -33,15 +33,21 @@ Accuracy is computed from JMH auxiliary counters in the current report. The coun
| Lucene SnowballFilter | 2.837% | 2.558% | 3.346% | Lucene TokenFilter integration path around the Snowball algorithm. |
| Official Snowball direct | 2.837% | 2.558% | 3.346% | Official Snowball generated Java stemmer; rule-based suffix algorithm. |
## Speed
Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fork, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
Speed uses JMH average time, 5 warmup iterations, 10 measurement iterations, 3 independent forks, and 1 thread. Relative factor is computed against the single Radixor row on this language page. Values below 1.000 are faster than that Radixor baseline; values above 1.000 are slower.
| Stemmer | Benchmark method | Score ms/op | Error ms | ns/token | Relative vs Radixor | Note |
| --- | --- | ---: | ---: | ---: | ---: | --- |
| Radixor | `radixor[YIDDISH]` | 0.254 | 0.004 | 50.7 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Official Snowball direct | `snowballDirect[YIDDISH]` | 1.537 | 0.220 | 307.3 | 6.058 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[YIDDISH]` | 1.714 | 0.120 | 342.8 | 6.756 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
| Radixor | `radixor[YIDDISH]` | 0.249 | 0.001 | 89.6 | 1.000 | Full Radixor dictionary patch-command stemmer. |
| Official Snowball direct | `snowballDirect[YIDDISH]` | 1.574 | 0.066 | 567.1 | 6.330 | Official Snowball generated Java stemmer; direct API. |
| Lucene SnowballFilter | `luceneSnowballFilter[YIDDISH]` | 1.849 | 0.079 | 665.9 | 7.434 | Lucene TokenFilter path around Snowball; includes TokenStream overhead. |
## Interpretation Notes
@@ -55,29 +61,29 @@ Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 1 fo
## Stemming Quality
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `YI` using the complete validated stemming-quality result matrix. Every usable dictionary row is one gold-standard group of forms expected to share a morphological family or lemma. Exact equality with a predetermined lemma is not required. Same-row pairs are positive pairs; pairs from different rows are negative pairs.
Runtime performance and linguistic grouping quality are independent dimensions. This section evaluates language `YI` using the complete validated stemming-quality result matrix. Every distinct surface form is one evaluated item and can belong to several dictionary groups. Two forms are a positive pair when their group-membership sets intersect and a negative pair when those sets are disjoint. A pair shared through several groups is counted once. Exact equality with a predetermined lemma is not required.
`ALL_WORDS` includes every valid group and its original forms. `LOWERCASE_GROUPS_ONLY` excludes an entire group when any Unicode code point is uppercase or titlecase; retained words are not lowercased or otherwise rewritten. This isolates case-handling effects without changing retained inputs. [Download the complete machine-readable result snapshot](../data/stemming-quality.csv).
### Evaluation Scope and Key Findings
The dictionary resource is `src/main/resources/yi/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
The default model is `yi-default`, loaded from classpath resource `org/egothor/stemmer/models/yi-default/stemmer.gz`. The following findings compare only deterministic `PRIMARY_OUTPUT` rows over identical included groups; candidate policies are reported separately as capability analyses.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.988241** among 3 deterministic stemmers. The runner-up is `SNOWBALL YIDDISH DIRECT` at 0.890988, a difference of 0.097253. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.988241** among 3 deterministic stemmers. The runner-up is `SNOWBALL YIDDISH DIRECT` at 0.890988, a difference of 0.097253. This rank does not imply leadership in throughput or every secondary metric.
- **ALL_WORDS:** `Radixor` ranks first by balanced accuracy at **0.989079** among 3 deterministic stemmers. The runner-up is `SNOWBALL YIDDISH DIRECT` at 0.891118, a difference of 0.097961. This rank does not imply leadership in throughput or every secondary metric.
- **LOWERCASE_GROUPS_ONLY:** `Radixor` ranks first by balanced accuracy at **0.989079** among 3 deterministic stemmers. The runner-up is `SNOWBALL YIDDISH DIRECT` at 0.891118, a difference of 0.097961. This rank does not imply leadership in throughput or every secondary metric.
### `ALL_WORDS`
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988241|195 / 6392909 (0.003050%)|149 / 6344 (2.348676%)|0.970881|0.972986|0.972965|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.890988|1151 / 6392909 (0.018004%)|1382 / 6344 (21.784363%)|0.805624|0.796661|0.796600|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.890988|1151 / 6392909 (0.018004%)|1382 / 6344 (21.784363%)|0.805624|0.796661|0.796600|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.989079|0.000000%|2.184236%|
|2|SNOWBALL YIDDISH DIRECT|0.891118|0.013211%|21.763216%|
|3|SNOWBALL YIDDISH LUCENE FILTER|0.891118|0.013211%|21.763216%|
</div>
@@ -85,9 +91,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.969484|0.976513|0.999969|0.988241|0.999946|0.000054|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.811713|0.782156|0.999820|0.890988|0.999604|0.000396|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.811713|0.782156|0.999820|0.890988|0.999604|0.000396|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.978158|1.000000|0.989079|0.999978|0.000022|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.857267|0.782368|0.999868|0.891118|0.999648|0.000352|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.857267|0.782368|0.999868|0.891118|0.999648|0.000352|
</details>
@@ -95,19 +101,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.970881|0.972986|0.975099|0.947393|0.972992|0.972965|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.805624|0.796661|0.787894|0.662041|0.796798|0.796600|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.805624|0.796661|0.787894|0.662041|0.796798|0.796600|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.972959|0.995691|0.996142|0.995917|0.995917|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.796462|0.982919|0.962014|0.972354|0.972354|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.796462|0.982919|0.962014|0.972354|0.972354|
|1|Radixor|PRIMARY_OUTPUT|0.995554|0.988958|0.982449|0.978158|0.989019|0.989008|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.841161|0.818107|0.796282|0.692200|0.818961|0.818787|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.841161|0.818107|0.796282|0.692200|0.818961|0.818787|
</details>
@@ -115,61 +111,39 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|6195|195|149|6392714|195 / 6392909|149 / 6344|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|4962|1151|1382|6391758|1151 / 6392909|1382 / 6344|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|4962|1151|1382|6391758|1151 / 6392909|1382 / 6344|
|1|Radixor|PRIMARY_OUTPUT|6180|0|138|6229428|0 / 6229428|138 / 6318|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|4943|823|1375|6228605|823 / 6229428|1375 / 6318|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|4943|823|1375|6228605|823 / 6229428|1375 / 6318|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 6392909 (0.000000%)|0 / 6344 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|6344|0|0|6392909|0 / 6392909|0 / 6344|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 6229428|0 / 6318|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999970|389 / 6392909 (0.006085%)|0 / 6344 (0.000000%)|0.953240|0.970253|0.970653|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -177,7 +151,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.942225|1.000000|0.999939|0.999970|0.999939|0.000061|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -185,15 +159,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.953240|0.970253|0.987885|0.942225|0.970683|0.970653|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -201,7 +167,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|6344|389|0|6392520|389 / 6392909|0 / 6344|
|1|Radixor|ALL_CANDIDATES|6318|0|0|6229428|0 / 6229428|0 / 6318|
</details>
@@ -211,21 +177,21 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|149|195|194|89|2.487423%|3|3676|
|Radixor|138|0|0|43|1.217441%|3|3578|
### `LOWERCASE_GROUPS_ONLY`
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. Rankings are separated by output policy and ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output policies**. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. `PRIMARY_OUTPUT` and `ALL_CANDIDATES` rankings are ordered by unrounded balanced accuracy, followed by MCC, F1, over-stemming rate, over-stemming count, under-stemming rate, and stemmer. `ANY_CANDIDATE` has no single rank metric and is listed alphabetically. Balanced accuracy is a navigation metric, not a universally authoritative quality score.
#### `PRIMARY_OUTPUT` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.988241|195 / 6392909 (0.003050%)|149 / 6344 (2.348676%)|0.970881|0.972986|0.972965|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.890988|1151 / 6392909 (0.018004%)|1382 / 6344 (21.784363%)|0.805624|0.796661|0.796600|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.890988|1151 / 6392909 (0.018004%)|1382 / 6344 (21.784363%)|0.805624|0.796661|0.796600|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|0.989079|0.000000%|2.184236%|
|2|SNOWBALL YIDDISH DIRECT|0.891118|0.013211%|21.763216%|
|3|SNOWBALL YIDDISH LUCENE FILTER|0.891118|0.013211%|21.763216%|
</div>
@@ -233,9 +199,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.969484|0.976513|0.999969|0.988241|0.999946|0.000054|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.811713|0.782156|0.999820|0.890988|0.999604|0.000396|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.811713|0.782156|0.999820|0.890988|0.999604|0.000396|
|1|Radixor|PRIMARY_OUTPUT|1.000000|0.978158|1.000000|0.989079|0.999978|0.000022|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.857267|0.782368|0.999868|0.891118|0.999648|0.000352|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.857267|0.782368|0.999868|0.891118|0.999648|0.000352|
</details>
@@ -243,19 +209,9 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.970881|0.972986|0.975099|0.947393|0.972992|0.972965|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.805624|0.796661|0.787894|0.662041|0.796798|0.796600|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.805624|0.796661|0.787894|0.662041|0.796798|0.796600|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|0.972959|0.995691|0.996142|0.995917|0.995917|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.796462|0.982919|0.962014|0.972354|0.972354|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.796462|0.982919|0.962014|0.972354|0.972354|
|1|Radixor|PRIMARY_OUTPUT|0.995554|0.988958|0.982449|0.978158|0.989019|0.989008|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|0.841161|0.818107|0.796282|0.692200|0.818961|0.818787|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|0.841161|0.818107|0.796282|0.692200|0.818961|0.818787|
</details>
@@ -263,61 +219,39 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|PRIMARY_OUTPUT|6195|195|149|6392714|195 / 6392909|149 / 6344|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|4962|1151|1382|6391758|1151 / 6392909|1382 / 6344|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|4962|1151|1382|6391758|1151 / 6392909|1382 / 6344|
|1|Radixor|PRIMARY_OUTPUT|6180|0|138|6229428|0 / 6229428|138 / 6318|
|2|SNOWBALL YIDDISH DIRECT|PRIMARY_OUTPUT|4943|823|1375|6228605|823 / 6229428|1375 / 6318|
|3|SNOWBALL YIDDISH LUCENE FILTER|PRIMARY_OUTPUT|4943|823|1375|6228605|823 / 6229428|1375 / 6318|
</details>
#### `ANY_CANDIDATE` ranking
#### `ANY_CANDIDATE` oracle bounds
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
These results are measured, not missing. `ANY_CANDIDATE` answers two separate optimistic questions for each pair: a gold-related pair avoids under-stemming when the candidate sets intersect, while a gold-negative pair avoids over-stemming when some non-colliding candidate selection exists. The oracle may choose a different candidate for the same word in different pairs. Consequently, these decisions do not form one globally realizable predicted relation or one TP/FP/FN/TN confusion matrix. Balanced accuracy, F-scores, Jaccard, Fowlkes–Mallows, and MCC are therefore mathematically **not applicable**, rather than unknown.
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|0 / 6392909 (0.000000%)|0 / 6344 (0.000000%)|1.000000|1.000000|1.000000|
<div class="quality-summary quality-summary--oracle" markdown="1">
| Stemmer | Optimistic over-stemming (OI) | Optimistic under-stemming (UI) |
|---|---:|---:|
|Radixor|0.000000%|0.000000%|
</div>
<details class="quality-details" markdown="1"><summary>Classification metrics</summary>
<details class="quality-details" markdown="1"><summary>Oracle-bound pair counts</summary>
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
<details class="quality-details" markdown="1"><summary>Pair-relation metrics</summary>
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|n/a|n/a|n/a|n/a|n/a|
</details>
<details class="quality-details" markdown="1"><summary>Raw pair counts</summary>
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ANY_CANDIDATE|6344|0|0|6392909|0 / 6392909|0 / 6344|
| Stemmer | Unavoidable over errors / gold-negative pairs | Unrepairable under errors / gold-related pairs |
|---|---:|---:|
|Radixor|0 / 6229428|0 / 6318|
</details>
#### `ALL_CANDIDATES` ranking
<div class="quality-table quality-table--compact" role="region" aria-label="Compact stemming-quality ranking; scroll horizontally for additional columns" tabindex="0" markdown="1">
<div class="quality-summary" markdown="1">
| Rank | Stemmer | Output policy | Balanced accuracy | Over-stemming | Under-stemming | F0.5 | F1 | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.999970|389 / 6392909 (0.006085%)|0 / 6344 (0.000000%)|0.953240|0.970253|0.970653|
| Rank | Stemmer | Balanced accuracy | Over-stemming (OI) | Under-stemming (UI) |
|---:|---|---:|---:|---:|
|1|Radixor|1.000000|0.000000%|0.000000%|
</div>
@@ -325,7 +259,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | Precision | Recall | Specificity | Balanced accuracy | Pairwise accuracy | Error rate |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.942225|1.000000|0.999939|0.999970|0.999939|0.000061|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|0.000000|
</details>
@@ -333,15 +267,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | F0.5 | F1 | F2 | Jaccard | Fowlkes–Mallows | MCC |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|0.953240|0.970253|0.987885|0.942225|0.970683|0.970653|
</details>
<details class="quality-details" markdown="1"><summary>Partition metrics (PRIMARY_OUTPUT only)</summary>
| Rank | Stemmer | Output policy | Adjusted Rand Index | Homogeneity | Completeness | V-measure | Normalized mutual information |
|---:|---|---|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|n/a|n/a|n/a|n/a|n/a|
|1|Radixor|ALL_CANDIDATES|1.000000|1.000000|1.000000|1.000000|1.000000|1.000000|
</details>
@@ -349,7 +275,7 @@ This mode contains **5 result rows**, **3 evaluated stemmers**, and **3 output p
| Rank | Stemmer | Output policy | TP | FP | FN | TN | Over error / possible | Under error / possible |
|---:|---|---|---:|---:|---:|---:|---:|---:|
|1|Radixor|ALL_CANDIDATES|6344|389|0|6392520|389 / 6392909|0 / 6344|
|1|Radixor|ALL_CANDIDATES|6318|0|0|6229428|0 / 6229428|0 / 6318|
</details>
@@ -359,19 +285,19 @@ Alternative candidates are capability analyses, not replacements for the determi
| Stemmer | Under pairs repaired | Best-case over pairs avoided | All-candidate collisions added | Multi-candidate forms | Multi-candidate share | Maximum candidates | Total candidate assignments |
|---|---:|---:|---:|---:|---:|---:|---:|
|Radixor|149|195|194|89|2.487423%|3|3676|
|Radixor|138|0|0|43|1.217441%|3|3578|
### Output Policies and Metric Definitions
`PRIMARY_OUTPUT` uses one deterministic stem per form and therefore defines a strict partition. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a same-group pair succeeds when candidates intersect, while a different-group pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and need not represent one globally consistent assignment. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect. Alternatives can reduce under-stemming but can introduce cross-group collisions, and the resulting relation can overlap and need not be a partition.
Each distinct surface form is one item and may belong to several gold groups. Two forms are gold-related when their membership sets intersect; a relation shared by several groups is counted once. `PRIMARY_OUTPUT` uses one deterministic stem per form. `ANY_CANDIDATE` is an optimistic oracle-assisted pairwise upper bound: a gold-related pair succeeds when candidates intersect, while a gold-negative pair succeeds when a non-colliding selection exists. Candidate choices may differ between pairs, so this is not deterministic runtime behaviour and does not define one confusion matrix. `ALL_CANDIDATES` activates every returned candidate; forms are related when candidate sets intersect.
For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. TP and FN concern same-group pairs; FP and TN concern different-group pairs. Consequently, under-stemming and over-stemming use different denominators. Undefined values are rendered as `n/a`.
For `PRIMARY_OUTPUT` and `ALL_CANDIDATES`, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs`, `FP = overErrorPairs`, and `TN = overPossiblePairs - overErrorPairs`. `ANY_CANDIDATE` publishes only its separate oracle-assisted under/over bounds; confusion-derived metrics are mathematically inapplicable and are not presented in its language-page section. Their machine-readable CSV fields remain empty. Undefined metric denominators in otherwise applicable policies are rendered as `n/a`.
- Under-stemming rate: `FN / (TP + FN)`, the false-negative rate over same-group pairs.
- Over-stemming rate: `FP / (TN + FP)`, the false-positive rate over different-group pairs.
- Under-stemming rate (Paice UI): `FN / (TP + FN)`, the false-negative rate over gold-related pairs.
- Over-stemming rate (Paice OI): `FP / (TN + FP)`, the false-positive rate over gold-negative pairs.
- Pairwise precision: `TP / (TP + FP)`, the fraction of predicted conflations that are gold-standard positive pairs.
- Pairwise recall: `TP / (TP + FN)`, the fraction of gold-standard positive pairs successfully connected.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of different-group pairs correctly separated.
- Pairwise specificity: `TN / (TN + FP)`, the fraction of gold-negative pairs correctly separated.
- Balanced accuracy: `(recall + specificity) / 2`. It gives equal weight to positive and negative pair classes and is less dominated by the large true-negative class than ordinary accuracy. It does not replace the raw errors or other metrics.
- Pairwise F-beta: `((1 + betaSquared) * TP) / (((1 + betaSquared) * TP) + (betaSquared * FN) + FP)`. F0.5 emphasizes precision and penalizes over-stemming more; F1 weights precision and recall equally; F2 emphasizes recall and penalizes under-stemming more.
- MCC: `(TP * TN - FP * FN) / sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN))`. It uses all confusion counts and remains useful under class imbalance, except when its denominator is degenerate.
@@ -380,16 +306,17 @@ For each row, `TP = underPossiblePairs - underErrorPairs`, `FN = underErrorPairs
- Pairwise accuracy: `(TP + TN) / (TP + TN + FP + FN)`. It can be dominated by true-negative cross-group pairs.
- Pairwise error rate: `(FP + FN) / (TP + TN + FP + FN)`.
Adjusted Rand Index uses the gold/predicted contingency table and chance correction. Homogeneity is `1 - H(gold | predicted) / H(gold)`; completeness is `1 - H(predicted | gold) / H(predicted)`; V-measure is their harmonic mean; normalized mutual information uses the arithmetic-mean entropy normalization `MI / ((H(gold) + H(predicted)) / 2)`. These partition-only metrics apply to `PRIMARY_OUTPUT`; candidate-relation rows show `n/a`.
Standard ARI, homogeneity, completeness, V-measure, and NMI are not calculated: their usual contingency-table definitions require an exclusive gold partition, while this gold standard is an overlapping cover.
### Provenance
- Authoritative source: `docs/benchmarks/data/stemming-quality.csv`
- Source SHA-256: `5a93a6ab60e46489737cd649eb1ac48182114b9038f7f20195ab9d1c1fc0dd28`
- Evaluation command: `./gradlew stemmingQuality`
- Source SHA-256: `edf16b07be8a535943ddf37caeb8807755c95e9e1fb13244145f28be74b491d8`
- Evaluation command: `./gradlew stemmingQuality --no-daemon`
- Dictionary language: `YI`
- Processing modes: `ALL_WORDS`, `LOWERCASE_GROUPS_ONLY`
- Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and `gradle.lockfile`
- Radixor version, Git revision, generation date, JDK version, operating system, and dictionary revision: not recorded in the authoritative CSV
- Model ID, version, and SHA-256: recorded in every CSV row
- Run date, core source state, JDK, operating system, and hardware: recorded on the [benchmark environment page](../reference/environment.md)
<!-- STEMMING-QUALITY:END -->