# 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 -->