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.

Dictionary size: ★★★★☆. The exact count is 57,658 distinct usable word forms after parser-compatible filtering and exact, case-preserved deduplication. Stars rank dictionary size relative to all benchmarked dictionaries in five nearly equal groups; they do not measure linguistic quality or benchmark accuracy.

All speed values are environment-specific and were measured on the hardware and JVM listed in the benchmark overview. The command distribution, exact-root accuracy, and speed tables belong to the published 2026-09-11 Radixor/Java 4.4.0 snapshot. Speed benchmark operations process changed tokens. Accuracy uses the complete Radixor dictionary for the language.

How to read this page

Start with the corpus and patch-command distribution, then compare exact-root agreement with runtime. The dictionary-family experiment, edit-cost experiment, and pairwise linguistic evaluation answer separate questions. Their 10–90% curves use independent frozen protocols and must not be substituted for one another.

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	Distinct usable forms	Complete quality tokens	Already-root tokens	Changed tokens	Timing workload	JMH timing tokens
he-il-default	1.0.0	HE_IL	2,358	57,658	61,071	4,715	56,356	changed tokens	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
AppendCharacterCommand	Appends one character to the end of the word form.	3	0.005%
BackwardCompoundCommand	Applies a multi-step backward patch made from skip, delete, insert, and replace operations.	46,876	76.757%
DeleteSuffixCommand	Deletes one or more trailing characters from the word form.	9,059	14.834%
PreserveCommand	Returns the word form unchanged because it already matches the preferred root.	4,869	7.973%
ReplaceLastCharacterCommand	Replaces the final character of the word form.	264	0.432%

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	98.250%	98.149%	99.449%	Full default-model Radixor dictionary patch-command stemmer.

Speed

Speed uses JMH average time, 3 warmup iterations, 5 measurement iterations, 3 independent forks, and 1 thread.

Stemmer	Benchmark method	Score ms/op	Error ms	ns/token	Relative vs Radixor	Note
Radixor	hebrewRadixor	5.000	0.095	88.7	1.000	Full default-model Radixor dictionary patch-command stemmer.

Interpretation notes

• Radixor is a dictionary-trained patch-command stemmer. Its learned transformations can generalize beyond the word forms listed in the training resource.
• Traversal is part of the measured model configuration and is recorded with the benchmark provenance.
• Results are environment-specific and should be compared only with rows from the same benchmark run.

Dictionary-family generalization conclusion

This is the language-specific conclusion from the independent radixor-generalization-v1 baseline experiment. It is intentionally separate from the wider edit-cost protocol below; values from the two frozen snapshots are not substituted for one another.

Evidence

Model he-il-default version 1.0.0 is evaluated over five predeclared nested splits. Unseen metrics remove withheld occurrences whose normalized surface also appeared in training. Parentheses show the observed split minimum–maximum.

Training rows	Median unseen occurrences	Unseen all exact	Unseen changed exact	Unseen root preserved
10%	54,723	16.532% (15.398–18.499)	10.423% (9.169–12.678)	89.756% (88.191–91.158)
20%	48,464	17.734% (17.573–18.304)	11.846% (11.656–12.416)	89.000% (88.647–89.233)
30%	42,248	19.125% (18.459–19.270)	13.345% (12.730–13.462)	88.961% (87.481–90.603)
40%	36,073	19.758% (19.628–20.304)	14.068% (13.841–14.608)	89.236% (88.497–90.062)
50%	29,985	20.205% (19.833–20.807)	14.533% (14.185–15.123)	88.771% (87.642–90.375)
60%	23,856	20.530% (20.267–21.311)	15.043% (14.721–15.824)	88.702% (87.417–89.456)
70%	17,846	21.488% (20.548–21.734)	16.059% (15.072–16.256)	88.713% (88.166–89.569)
80%	11,868	21.715% (19.936–22.602)	16.441% (14.182–17.242)	89.628% (87.585–91.818)
90%	5,914	21.813% (21.558–23.195)	16.362% (15.846–17.917)	88.785% (87.416–93.636)

Generalization conclusion

• Median exactness on genuinely unseen changed forms moves from 10.423% at 10% training knowledge to 16.362% at 90%, a measured +5.939 percentage-point change.
• Unseen all-form exactness moves from 16.532% at 10% training knowledge to 21.813% at 90%, a measured +5.281 percentage-point change.
• Preservation of unseen already-root forms moves from 89.756% at 10% training knowledge to 88.785% at 90%, a measured -0.971 percentage-point change.
• The evidence establishes within-resource transfer across withheld dictionary families. It does not estimate unrelated domains, misspellings, arbitrary compounds, or external corpora.

The complete ten-level table and split ranges remain in the complete generalization appendix; raw counters and provenance are in active machine-readable snapshot. The frozen methodology defines family-level splitting, unseen-surface leakage control, aggregation, and the limits of the claim.

Edit costs and dictionary-knowledge generalization

This section interprets the edit-cost and held-out-family experiment for HE_IL separately from the cross-language macro summary. Each knowledge point is the median of five frozen, nested splits. The primary exactness outcome covers changed forms in withheld families after excluding normalized surfaces seen in training. Thus the complete dictionary is the evaluation population, while only genuinely unseen surfaces contribute to this outcome.

Cost labels have the fixed form D<delete>I<insert>R<replace>M<match>. D is the cost of deleting a source character, I of inserting a target character, R of replacing a source character, and M of keeping an equal source/target character unchanged (the match or skip step). For example, D2I5R3M0 means delete cost 2, insert cost 5, replace cost 3, and match cost 0. The numbers are relative dynamic-programming costs, not command counts.

Evidence

Dictionary rows	Evaluated forms	Changed-form share	Baseline commands	Exact cost classes	Grid reduction	Largest exact class
2,358	61,071	92.28%	4,150	13	18.00×	59

The exact classes are based on command-by-command equality over the complete dictionary, not equality of aggregate trie metrics. A higher class count means that this dictionary exposes more cost-dependent encoder decisions; it does not by itself mean better quality.

Knowledge	Baseline unseen changed exact	Selected-cost exact	Δ	Baseline F0.5	Selected F0.5	Baseline commands	Selected commands
10%	10.977%	10.991%	+0.014 pp	0.2241	0.2251	1.000×	0.991×
20%	11.977%	12.022%	+0.045 pp	0.2621	0.2634	1.000×	0.987×
30%	13.651%	13.692%	+0.041 pp	0.2771	0.2777	1.000×	0.979×
40%	14.083%	14.095%	+0.012 pp	0.2845	0.2849	1.000×	0.977×
50%	14.304%	14.340%	+0.036 pp	0.2871	0.2881	1.000×	0.973×
60%	14.804%	14.818%	+0.014 pp	0.3016	0.3018	1.000×	0.974×
70%	15.780%	15.852%	+0.073 pp	0.3057	0.3082	1.000×	0.974×
80%	15.425%	15.507%	+0.082 pp	0.3148	0.3184	1.000×	0.970×
90%	15.910%	16.074%	+0.164 pp	0.3374	0.3416	1.000×	0.970×

Within-language associations

Spearman coefficients are calculated independently inside each seed × knowledge stratum across the normalized cost grid. The table reports the median and central 95% empirical interval across up to 45 strata. A relationship is called stable only when it is defined in all 45 strata and the interval retains one sign. These intervals are descriptive, not multiplicity-adjusted confidence intervals. Every predictor and outcome label is defined in the methodology glossary.

The strongest structural pairs whose central interval retains one sign are:

Predictor	Structural outcome	Median Spearman ρ	Central 95%	Strata
patch_command_ratio	value_references	+1.000	+0.960…+1.000	45
patch_command_ratio	trie_nodes	+0.996	+0.907…+0.999	45
replace_to_delete_insert	patch_command_ratio	-0.825	-0.827…-0.823	45
replace_to_delete_insert	value_references	-0.825	-0.827…-0.783	45
replace_to_delete_insert	trie_nodes	-0.823	-0.826…-0.709	45
replace_cost	patch_command_ratio	-0.722	-0.727…-0.709	45

For each quality outcome, the largest absolute median association is shown even when its interval crosses zero. This prevents a large median in heterogeneous strata from being misreported as a portable language-level effect.

Predictor	Quality outcome	Median Spearman ρ	Central 95%	Stable	Defined strata
patch_command_ratio	unseen_changed_exact	-0.905	-0.953…-0.230	yes	45 / 45
average_path_length	unseen_f05	-0.937	-0.984…+0.524	no	45 / 45
replace_to_delete_insert	unseen_over_percent	-0.890	-0.934…-0.773	yes	45 / 45
average_path_length	unseen_under_percent	+0.913	-0.524…+0.984	no	45 / 45

Edit-cost conclusion

• With baseline costs, median unseen changed-form exactness changes from 10.977% at 10% knowledge to 15.910% at 90%, a +4.933 pp measured knowledge effect.
• The predeclared selection is D10I5R10M1. Its median unseen changed-form exactness differs from baseline by +0.022 pp and it reduces the median retained-command count by 2.48% (0.975× baseline).
• Under the selected costs, the 10%–90% knowledge change is +5.083 pp. This quantifies generalization for this dictionary; it is not a claim about unrelated domains or lexical resources.
• The non-baseline setting is an efficiency candidate, not a production default: it was selected and evaluated on the same matrix and therefore requires external-corpus or external-dictionary validation before adoption.
• 32 cost/representation-to-quality association(s) are defined in all 45 strata and retain one sign over their central 95% interval. Their direction is evidence for this resource only; inspect the table and machine-readable coefficients before extrapolating.

The complete evidence is available in the raw logical matrix, the per-language knowledge curves, and the per-language association table. See the cross-language analysis and frozen methodology for scope and limitations.

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 and the canonical linguistic-quality methodology, including its overlapping-group relation, dictionary modes, output policies, metrics, and ranking rules.

Radixor's model was trained from the same lexical resource that defines this benchmark's dictionary-reference relation. The result therefore measures same-resource agreement, not independent external linguistic validity. Third-party adapters receive the same evaluated forms but were not trained by this benchmark.

The tables retain both canonical dictionary-processing modes and every applicable output policy without redefining them on each language page. Download the complete machine-readable result snapshot.

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 has balanced accuracy 0.986253. No same-language comparative ranking is available because it is the only evaluated deterministic stemmer. This result does not imply leadership in throughput or every secondary metric.
• LOWERCASE_GROUPS_ONLY: Radixor has balanced accuracy 0.986253. No same-language comparative ranking is available because it is the only evaluated deterministic stemmer. This result does not imply leadership in throughput or every secondary metric.

ALL_WORDS

This mode contains 3 result rows, 1 evaluated stemmer, 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. Where at least two deterministic rows are available, 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 result (no same-language comparator)

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
n/a	Radixor	0.986253	0.000000%	2.749465%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
n/a	Radixor	PRIMARY_OUTPUT	1.000000	0.972505	1.000000	0.986253	0.999988	0.000012

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
n/a	Radixor	PRIMARY_OUTPUT	0.994377	0.986061	0.977883	0.972505	0.986157	0.986151

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
n/a	Radixor	PRIMARY_OUTPUT	686015	0	19395	1661488243	0 / 1661488243	19395 / 705410

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.

Stemmer	Optimistic over-stemming (OI)	Optimistic under-stemming (UI)
Radixor	0.000000%	0.000000%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 1661488243	0 / 705410

ALL_CANDIDATES result (no same-language comparator)

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
n/a	Radixor	1.000000	0.000000%	0.000000%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
n/a	Radixor	ALL_CANDIDATES	1.000000	1.000000	1.000000	1.000000	1.000000	0.000000

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
n/a	Radixor	ALL_CANDIDATES	1.000000	1.000000	1.000000	1.000000	1.000000	1.000000

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
n/a	Radixor	ALL_CANDIDATES	705410	0	0	1661488243	0 / 1661488243	0 / 705410

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	19395	0	0	984	1.706615%	40	58714

LOWERCASE_GROUPS_ONLY

This mode contains 3 result rows, 1 evaluated stemmer, 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. Where at least two deterministic rows are available, 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 result (no same-language comparator)

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
n/a	Radixor	0.986253	0.000000%	2.749465%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
n/a	Radixor	PRIMARY_OUTPUT	1.000000	0.972505	1.000000	0.986253	0.999988	0.000012

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
n/a	Radixor	PRIMARY_OUTPUT	0.994377	0.986061	0.977883	0.972505	0.986157	0.986151

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
n/a	Radixor	PRIMARY_OUTPUT	686015	0	19395	1661488243	0 / 1661488243	19395 / 705410

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.

Stemmer	Optimistic over-stemming (OI)	Optimistic under-stemming (UI)
Radixor	0.000000%	0.000000%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 1661488243	0 / 705410

ALL_CANDIDATES result (no same-language comparator)

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
n/a	Radixor	1.000000	0.000000%	0.000000%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
n/a	Radixor	ALL_CANDIDATES	1.000000	1.000000	1.000000	1.000000	1.000000	0.000000

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
n/a	Radixor	ALL_CANDIDATES	1.000000	1.000000	1.000000	1.000000	1.000000	1.000000

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
n/a	Radixor	ALL_CANDIDATES	705410	0	0	1661488243	0 / 1661488243	0 / 705410

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	19395	0	0	984	1.706615%	40	58714

Method and interpretation boundaries

The linguistic-quality methodology defines the overlapping gold relation, both dictionary-processing modes, all three output policies, confusion counts, formulas, undefined denominators, aggregation, ranking, and inapplicable partition metrics. The candidate-policy reference explains why ANY_CANDIDATE is an oracle-assisted capability bound rather than deterministic runtime behavior.

This page preserves the language-specific raw counts, metrics, candidate distributions, comparison availability, and caveats. The machine-readable CSV remains authoritative for every field.

Provenance

• Authoritative source: docs/benchmarks/data/stemming-quality-2026-09-13.csv
• Source SHA-256: d41e00160cda44758e806c37d210f2d9b90b1ebad22eebe5ece89a40f4ed9ab0
• 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