English stemmer benchmarks ★★★★★

This page reports same-language stemming benchmarks for English. Accuracy is listed first because speed without root agreement is not enough to interpret stemmer quality.

Dictionary size: ★★★★★. The exact count is 591,946 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.

Runtime and exact-root agreement measure different properties. Light, minimal, possessive, and other rule-based filters intentionally have different transformation scopes, so a lower runtime can coexist with lower dictionary-root agreement. Read the speed and accuracy tables together. The Radixor rows in this refresh use the contracted compiled patch trie: compilation collapses uniform patch-command subtrees into accepting leaves, reducing hot lookup depth while preserving the preferred stemming result measured by the accuracy pass. The EnglishRadixorDictionaryCoverageBenchmark shows the resulting quality/speed envelope explicitly.

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
us-uk-default	1.0.1	US_UK	396,939	591,946	1,002,414	793,874	208,540	changed tokens	208,540

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 1,002,414.

Command class	Meaning	Word forms	Share
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	2.243%
DeleteSuffixCommand	Deletes one or more trailing characters from the word form.	202,637	20.215%
PreserveCommand	Returns the word form unchanged because it already matches the preferred root.	777,146	77.527%
ReplaceLastCharacterCommand	Replaces the final character of the word form.	77	0.008%

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	97.668%	98.110%	97.552%	Radixor dictionary-trained patch-command stemmer.
Lucene EnglishMinimalStemFilter	91.159%	65.801%	97.820%	Minimal English plural reduction, not a full stemmer.
Lucene KStemFilter	80.233%	77.328%	80.996%	Krovetz-style English stemming TokenFilter; broader than minimal suffix reducers.
Lucene HunspellStemFilter	80.400%	12.869%	98.139%	Benchmark-only English Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene EnglishPossessiveFilter	79.186%	0.003%	99.987%	Possessive-ending remover only, not a full stemmer.
Official Snowball Porter2 (Java)	40.425%	46.737%	38.767%	Porter2 rule-based suffix stemmer, distinct from original Porter.
Lucene PorterStemFilter	39.616%	46.636%	37.772%	Lucene TokenFilter path for Porter suffix rules; not dictionary-root equivalent.
Lucene PorterStemmer direct copy	39.616%	46.636%	37.772%	Direct Porter suffix-rule implementation generated under build for benchmark-only use.
OpenNLP PorterStemmer	39.616%	46.636%	37.772%	Apache OpenNLP Porter suffix-rule implementation.
Snowball original Porter	40.425%	46.737%	38.767%	Classic Porter rule-based suffix stemmer.
Paice/Husk Lancaster	28.110%	37.387%	25.673%	Aggressive Paice/Husk rule stemmer that often produces shorter stems.

Speed

Speed uses JMH average time, 3 warmup iterations, 5 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.

Separate English speed suites

The 131.3 ns/token Radixor value below is from the multilingual same-language comparison suite. The coverage experiment reports its own full-knowledge point from a separate benchmark method and run. Treat both as suite-specific estimates with their published uncertainty, not as interchangeable values.

Stemmer	Benchmark method	Score ms/op	Error ms	ns/token	Relative vs Radixor	Note
Radixor	radixor[us-uk-default]	27.372	1.939	131.3	1.000	Full dictionary patch-command stemmer using compiled patch commands.
Lucene EnglishPossessiveFilter	luceneEnglishPossessiveFilter	16.688	0.282	80.0	0.610	Possessive-ending remover only; not a full stemmer.
Lucene EnglishMinimalStemFilter	luceneEnglishMinimalStemFilter	18.111	0.205	86.8	0.662	Narrow plural reduction filter; not a full stemmer.
Lucene PorterStemmer direct copy	lucenePorterStemmerCopied	19.393	2.846	93.0	0.708	Benchmark-only generated copy of Lucene package-private Porter implementation.
OpenNLP PorterStemmer	opennlpPorterStemmer	21.062	4.477	101.0	0.769	Apache OpenNLP Porter implementation.
Snowball original Porter	snowballOriginalPorter	36.617	1.685	175.6	1.338	Classic Porter suffix-rule stemmer; historical English baseline, not a dictionary-equivalent stemmer.
Lucene PorterStemFilter	lucenePorterStemFilter	31.828	0.534	152.6	1.163	Lucene TokenFilter integration path for Porter; includes TokenStream overhead.
Lucene KStemFilter	luceneKStemFilter	44.695	1.697	214.3	1.633	Krovetz-style English TokenFilter; broader than minimal suffix filters.
Lucene HunspellStemFilter	luceneHunspellStemFilter	76.009	6.313	364.5	2.777	Benchmark-only English Hunspell comparison using the benchmark Hunspell corpus.
Official Snowball Porter2 (Java)	snowballEnglishPorter2	47.762	0.989	229.0	1.745	Porter2 suffix-rule stemmer, distinct from original Porter.
Paice/Husk Lancaster	paiceHuskLancaster	144.087	2.304	690.9	5.264	Aggressive rule-based English 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.
• Light, minimal, plural, and possessive filters are narrow baselines. They can be fast because they intentionally perform less linguistic work.
• Lucene TokenFilter rows include TokenStream, attribute, and required normalization overhead. Direct rows measure exposed direct APIs.
• Morfologik rows are dictionary-based and can emit multiple terms for one input token. Quality rows use the first returned term when no ranking weight is available.
• Snowball rows are rule-based generated suffix stemmers; they are useful algorithmic baselines, not dictionary-root equivalence guarantees.

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 us-uk-default version 1.0.1 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%	898,572	92.805% (92.757–92.843)	76.010% (75.505–76.378)	97.184% (97.156–97.293)
20%	795,287	93.155% (93.122–93.220)	76.608% (76.306–76.669)	97.509% (97.464–97.541)
30%	692,940	93.440% (93.409–93.459)	76.944% (76.715–76.961)	97.741% (97.718–97.755)
40%	591,318	93.699% (93.682–93.712)	77.308% (77.149–77.479)	97.940% (97.912–97.948)
50%	490,475	93.934% (93.888–94.011)	77.643% (77.410–77.726)	98.133% (98.089–98.192)
60%	390,750	94.178% (94.121–94.242)	77.976% (77.824–78.311)	98.314% (98.268–98.319)
70%	291,831	94.425% (94.290–94.438)	78.505% (77.971–78.614)	98.454% (98.436–98.500)
80%	193,781	94.608% (94.514–94.637)	78.795% (78.580–79.033)	98.592% (98.577–98.658)
90%	96,702	94.680% (94.609–94.847)	78.854% (78.538–79.603)	98.744% (98.642–98.790)

Generalization conclusion

• Median exactness on genuinely unseen changed forms moves from 76.010% at 10% training knowledge to 78.854% at 90%, a measured +2.843 percentage-point change.
• Unseen all-form exactness moves from 92.805% at 10% training knowledge to 94.680% at 90%, a measured +1.875 percentage-point change.
• Preservation of unseen already-root forms moves from 97.184% at 10% training knowledge to 98.744% at 90%, a measured +1.560 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 US_UK 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
396,939	1,002,414	20.80%	337	15	15.60×	54

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%	75.824%	75.824%	+0.000 pp	0.8800	0.8800	1.000×	1.000×
20%	76.283%	76.283%	+0.000 pp	0.8841	0.8841	1.000×	1.000×
30%	76.794%	76.794%	+0.000 pp	0.8890	0.8890	1.000×	1.000×
40%	77.289%	77.289%	+0.000 pp	0.8928	0.8928	1.000×	1.000×
50%	77.779%	77.779%	+0.000 pp	0.8961	0.8961	1.000×	1.000×
60%	78.057%	78.057%	+0.000 pp	0.8985	0.8985	1.000×	1.000×
70%	78.403%	78.403%	+0.000 pp	0.9020	0.9020	1.000×	1.000×
80%	78.806%	78.806%	+0.000 pp	0.9054	0.9054	1.000×	1.000×
90%	79.131%	79.131%	+0.000 pp	0.9076	0.9076	1.000×	1.000×

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	+0.815	+0.683…+1.000	45
replace_to_delete_insert	patch_command_ratio	-0.794	-0.885…-0.599	45
replace_cost	patch_command_ratio	-0.687	-0.737…-0.437	45
patch_command_ratio	trie_nodes	+0.807	+0.351…+1.000	45
replace_to_delete_insert	trie_nodes	-0.669	-0.794…-0.324	45
replace_to_delete_insert	value_references	-0.723	-0.805…-0.313	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
trie_edges	unseen_f05	+1.000	+0.250…+1.000	no	4 / 45
trie_edges	unseen_over_percent	-1.000	-1.000…-0.250	no	4 / 45

No within-stratum coefficient is defined for unseen_changed_exact, unseen_under_percent because these outcomes do not vary across cost configurations in the measured language strata. Within this matrix, that is observed cost insensitivity for those outcomes, not missing measurement.

Edit-cost conclusion

• With baseline costs, median unseen changed-form exactness changes from 75.824% at 10% knowledge to 79.131% at 90%, a +3.308 pp measured knowledge effect.
• The predeclared selection is D1I1R1M0. Its median unseen changed-form exactness differs from baseline by +0.000 pp and it does not change the median retained-command count (1.000× baseline).
• Under the selected costs, the 10%–90% knowledge change is +3.308 pp. This quantifies generalization for this dictionary; it is not a claim about unrelated domains or lexical resources.
• The selection rule retains the production baseline, so this experiment supplies no measured reason to change edit costs for this language under the predeclared objective.
• No cost or representation predictor is both defined in all 45 strata and retains one association sign over the central 95% interval for an unseen-form quality outcome. Effects with partial coverage are insufficient for a stable language-level claim; the remaining measured effects are heterogeneous across knowledge levels and splits.

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 US_UK 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 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 0.976120 among 11 deterministic stemmers. The runner-up is ENGLISH LUCENE PORTER COPIED at 0.965135, a difference of 0.010985. This rank does not imply leadership in throughput or every secondary metric.
• LOWERCASE_GROUPS_ONLY: Radixor ranks first by balanced accuracy at 0.976863 among 11 deterministic stemmers. The runner-up is ENGLISH LUCENE PORTER COPIED at 0.965469, a difference of 0.011393. 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. 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 ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	0.976120	<0.000001%	4.775939%
2	ENGLISH LUCENE PORTER COPIED	0.965135	0.000207%	6.972812%
3	ENGLISH LUCENE PORTER FILTER	0.965135	0.000207%	6.972812%
4	ENGLISH OPENNLP PORTER	0.965135	0.000207%	6.972812%
5	ENGLISH SNOWBALL PORTER2	0.965070	0.000212%	6.985868%
6	ENGLISH SNOWBALL ORIGINAL PORTER	0.964995	0.000206%	7.000881%
7	ENGLISH PAICE HUSK LANCASTER	0.962815	0.000960%	7.435948%
8	ENGLISH LUCENE KSTEM FILTER	0.887253	0.000110%	22.549365%
9	ENGLISH LUCENE MINIMAL FILTER	0.723935	0.000001%	55.212964%
10	HUNSPELL ENGLISH LUCENE FILTER	0.574800	0.000012%	85.039982%
11	ENGLISH LUCENE POSSESSIVE FILTER	0.500011	<0.000001%	99.997715%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	0.999990	0.952241	1.000000	0.976120	1.000000	0.000000
2	ENGLISH LUCENE PORTER COPIED	PRIMARY_OUTPUT	0.440121	0.930272	0.999998	0.965135	0.999998	0.000002
3	ENGLISH LUCENE PORTER FILTER	PRIMARY_OUTPUT	0.440121	0.930272	0.999998	0.965135	0.999998	0.000002
4	ENGLISH OPENNLP PORTER	PRIMARY_OUTPUT	0.440121	0.930272	0.999998	0.965135	0.999998	0.000002
5	ENGLISH SNOWBALL PORTER2	PRIMARY_OUTPUT	0.434309	0.930141	0.999998	0.965070	0.999998	0.000002
6	ENGLISH SNOWBALL ORIGINAL PORTER	PRIMARY_OUTPUT	0.441440	0.929991	0.999998	0.964995	0.999998	0.000002
7	ENGLISH PAICE HUSK LANCASTER	PRIMARY_OUTPUT	0.144282	0.925641	0.999990	0.962815	0.999990	0.000010
8	ENGLISH LUCENE KSTEM FILTER	PRIMARY_OUTPUT	0.551014	0.774506	0.999999	0.887253	0.999999	0.000001
9	ENGLISH LUCENE MINIMAL FILTER	PRIMARY_OUTPUT	0.989894	0.447870	1.000000	0.723935	0.999999	0.000001
10	HUNSPELL ENGLISH LUCENE FILTER	PRIMARY_OUTPUT	0.681262	0.149600	1.000000	0.574800	0.999998	0.000002
11	ENGLISH LUCENE POSSESSIVE FILTER	PRIMARY_OUTPUT	0.148936	0.000023	1.000000	0.500011	0.999998	0.000002

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.990061	0.975531	0.961422	0.952231	0.975823	0.975823
2	ENGLISH LUCENE PORTER COPIED	PRIMARY_OUTPUT	0.491963	0.597539	0.760812	0.426065	0.639869	0.639868
3	ENGLISH LUCENE PORTER FILTER	PRIMARY_OUTPUT	0.491963	0.597539	0.760812	0.426065	0.639869	0.639868
4	ENGLISH OPENNLP PORTER	PRIMARY_OUTPUT	0.491963	0.597539	0.760812	0.426065	0.639869	0.639868
5	ENGLISH SNOWBALL PORTER2	PRIMARY_OUTPUT	0.486138	0.592134	0.757239	0.420590	0.635585	0.635584
6	ENGLISH SNOWBALL ORIGINAL PORTER	PRIMARY_OUTPUT	0.493266	0.598696	0.761449	0.427243	0.640731	0.640730
7	ENGLISH PAICE HUSK LANCASTER	PRIMARY_OUTPUT	0.173588	0.249650	0.444357	0.142629	0.365449	0.365447
8	ENGLISH LUCENE KSTEM FILTER	PRIMARY_OUTPUT	0.584762	0.643919	0.716392	0.474838	0.653272	0.653271
9	ENGLISH LUCENE MINIMAL FILTER	PRIMARY_OUTPUT	0.796987	0.616713	0.502949	0.445832	0.665841	0.665840
10	HUNSPELL ENGLISH LUCENE FILTER	PRIMARY_OUTPUT	0.398218	0.245328	0.177269	0.139814	0.319244	0.319244
11	ENGLISH LUCENE POSSESSIVE FILTER	PRIMARY_OUTPUT	0.000114	0.000046	0.000029	0.000023	0.001845	0.001845

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	291757	3	14633	175199431092	3 / 175199431095	14633 / 306390
2	ENGLISH LUCENE PORTER COPIED	PRIMARY_OUTPUT	285026	362583	21364	175199068512	362583 / 175199431095	21364 / 306390
3	ENGLISH LUCENE PORTER FILTER	PRIMARY_OUTPUT	285026	362583	21364	175199068512	362583 / 175199431095	21364 / 306390
4	ENGLISH OPENNLP PORTER	PRIMARY_OUTPUT	285026	362583	21364	175199068512	362583 / 175199431095	21364 / 306390
5	ENGLISH SNOWBALL PORTER2	PRIMARY_OUTPUT	284986	371197	21404	175199059898	371197 / 175199431095	21404 / 306390
6	ENGLISH SNOWBALL ORIGINAL PORTER	PRIMARY_OUTPUT	284940	360538	21450	175199070557	360538 / 175199431095	21450 / 306390
7	ENGLISH PAICE HUSK LANCASTER	PRIMARY_OUTPUT	283607	1682038	22783	175197749057	1682038 / 175199431095	22783 / 306390
8	ENGLISH LUCENE KSTEM FILTER	PRIMARY_OUTPUT	237301	193361	69089	175199237734	193361 / 175199431095	69089 / 306390
9	ENGLISH LUCENE MINIMAL FILTER	PRIMARY_OUTPUT	137223	1401	169167	175199429694	1401 / 175199431095	169167 / 306390
10	HUNSPELL ENGLISH LUCENE FILTER	PRIMARY_OUTPUT	45836	21445	260554	175199409650	21445 / 175199431095	260554 / 306390
11	ENGLISH LUCENE POSSESSIVE FILTER	PRIMARY_OUTPUT	7	40	306383	175199431055	40 / 175199431095	306383 / 306390

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%	0.004569%
HUNSPELL ENGLISH LUCENE FILTER	0.000012%	83.349652%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 175199431095	14 / 306390
HUNSPELL ENGLISH LUCENE FILTER	20368 / 175199431095	255375 / 306390

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	0.999977	<0.000001%	0.004569%
2	HUNSPELL ENGLISH LUCENE FILTER	0.583252	0.000022%	83.349652%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	ALL_CANDIDATES	0.999827	0.999954	1.000000	0.999977	1.000000	0.000000
2	HUNSPELL ENGLISH LUCENE FILTER	ALL_CANDIDATES	0.568121	0.166503	1.000000	0.583252	0.999998	0.000002

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	ALL_CANDIDATES	0.999852	0.999891	0.999929	0.999781	0.999891	0.999891
2	HUNSPELL ENGLISH LUCENE FILTER	ALL_CANDIDATES	0.383241	0.257531	0.193921	0.147796	0.307562	0.307561

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	306376	53	14	175199431042	53 / 175199431095	14 / 306390
2	HUNSPELL ENGLISH LUCENE FILTER	ALL_CANDIDATES	51015	38781	255375	175199392314	38781 / 175199431095	255375 / 306390

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	14619	3	50	13716	2.317103%	8	605958
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. 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 ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	0.976863	<0.000001%	4.627473%
2	ENGLISH LUCENE PORTER COPIED	0.965469	0.000222%	6.905897%
3	ENGLISH LUCENE PORTER FILTER	0.965469	0.000222%	6.905897%
4	ENGLISH OPENNLP PORTER	0.965469	0.000222%	6.905897%
5	ENGLISH SNOWBALL PORTER2	0.965445	0.000228%	6.910824%
6	ENGLISH SNOWBALL ORIGINAL PORTER	0.965328	0.000221%	6.934147%
7	ENGLISH PAICE HUSK LANCASTER	0.963199	0.001032%	7.359216%
8	ENGLISH LUCENE KSTEM FILTER	0.889741	0.000120%	22.051698%
9	ENGLISH LUCENE MINIMAL FILTER	0.724902	0.000001%	55.019529%
10	HUNSPELL ENGLISH LUCENE FILTER	0.575162	0.000012%	84.967529%
11	ENGLISH LUCENE POSSESSIVE FILTER	0.500008	<0.000001%	99.998358%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	0.999990	0.953725	1.000000	0.976863	1.000000	0.000000
2	ENGLISH LUCENE PORTER COPIED	PRIMARY_OUTPUT	0.440920	0.930941	0.999998	0.965469	0.999998	0.000002
3	ENGLISH LUCENE PORTER FILTER	PRIMARY_OUTPUT	0.440920	0.930941	0.999998	0.965469	0.999998	0.000002
4	ENGLISH OPENNLP PORTER	PRIMARY_OUTPUT	0.440920	0.930941	0.999998	0.965469	0.999998	0.000002
5	ENGLISH SNOWBALL PORTER2	PRIMARY_OUTPUT	0.435153	0.930892	0.999998	0.965445	0.999998	0.000002
6	ENGLISH SNOWBALL ORIGINAL PORTER	PRIMARY_OUTPUT	0.442235	0.930659	0.999998	0.965328	0.999998	0.000002
7	ENGLISH PAICE HUSK LANCASTER	PRIMARY_OUTPUT	0.144698	0.926408	0.999990	0.963199	0.999990	0.000010
8	ENGLISH LUCENE KSTEM FILTER	PRIMARY_OUTPUT	0.551013	0.779483	0.999999	0.889741	0.999998	0.000002
9	ENGLISH LUCENE MINIMAL FILTER	PRIMARY_OUTPUT	0.989965	0.449805	1.000000	0.724902	0.999999	0.000001
10	HUNSPELL ENGLISH LUCENE FILTER	PRIMARY_OUTPUT	0.700121	0.150325	1.000000	0.575162	0.999998	0.000002
11	ENGLISH LUCENE POSSESSIVE FILTER	PRIMARY_OUTPUT	0.121951	0.000016	1.000000	0.500008	0.999998	0.000002

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.990381	0.976310	0.962632	0.953716	0.976584	0.976583
2	ENGLISH LUCENE PORTER COPIED	PRIMARY_OUTPUT	0.492799	0.598414	0.761648	0.426955	0.640680	0.640679
3	ENGLISH LUCENE PORTER FILTER	PRIMARY_OUTPUT	0.492799	0.598414	0.761648	0.426955	0.640680	0.640679
4	ENGLISH OPENNLP PORTER	PRIMARY_OUTPUT	0.492799	0.598414	0.761648	0.426955	0.640680	0.640679
5	ENGLISH SNOWBALL PORTER2	PRIMARY_OUTPUT	0.487025	0.593070	0.758150	0.421535	0.636459	0.636458
6	ENGLISH SNOWBALL ORIGINAL PORTER	PRIMARY_OUTPUT	0.494097	0.599565	0.762279	0.428128	0.641537	0.641536
7	ENGLISH PAICE HUSK LANCASTER	PRIMARY_OUTPUT	0.174075	0.250301	0.445287	0.143054	0.366127	0.366125
8	ENGLISH LUCENE KSTEM FILTER	PRIMARY_OUTPUT	0.585325	0.645632	0.719793	0.476703	0.655367	0.655366
9	ENGLISH LUCENE MINIMAL FILTER	PRIMARY_OUTPUT	0.798246	0.618559	0.504903	0.447763	0.667301	0.667301
10	HUNSPELL ENGLISH LUCENE FILTER	PRIMARY_OUTPUT	0.404349	0.247507	0.178333	0.141231	0.324416	0.324415
11	ENGLISH LUCENE POSSESSIVE FILTER	PRIMARY_OUTPUT	0.000082	0.000033	0.000021	0.000016	0.001415	0.001415

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	290334	3	14087	161561996596	3 / 161561996599	14087 / 304421
2	ENGLISH LUCENE PORTER COPIED	PRIMARY_OUTPUT	283398	359344	21023	161561637255	359344 / 161561996599	21023 / 304421
3	ENGLISH LUCENE PORTER FILTER	PRIMARY_OUTPUT	283398	359344	21023	161561637255	359344 / 161561996599	21023 / 304421
4	ENGLISH OPENNLP PORTER	PRIMARY_OUTPUT	283398	359344	21023	161561637255	359344 / 161561996599	21023 / 304421
5	ENGLISH SNOWBALL PORTER2	PRIMARY_OUTPUT	283383	367843	21038	161561628756	367843 / 161561996599	21038 / 304421
6	ENGLISH SNOWBALL ORIGINAL PORTER	PRIMARY_OUTPUT	283312	357325	21109	161561639274	357325 / 161561996599	21109 / 304421
7	ENGLISH PAICE HUSK LANCASTER	PRIMARY_OUTPUT	282018	1666994	22403	161560329605	1666994 / 161561996599	22403 / 304421
8	ENGLISH LUCENE KSTEM FILTER	PRIMARY_OUTPUT	237291	193354	67130	161561803245	193354 / 161561996599	67130 / 304421
9	ENGLISH LUCENE MINIMAL FILTER	PRIMARY_OUTPUT	136930	1388	167491	161561995211	1388 / 161561996599	167491 / 304421
10	HUNSPELL ENGLISH LUCENE FILTER	PRIMARY_OUTPUT	45762	19601	258659	161561976998	19601 / 161561996599	258659 / 304421
11	ENGLISH LUCENE POSSESSIVE FILTER	PRIMARY_OUTPUT	5	36	304416	161561996563	36 / 161561996599	304416 / 304421

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%
HUNSPELL ENGLISH LUCENE FILTER	0.000011%	83.267252%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 161561996599	0 / 304421
HUNSPELL ENGLISH LUCENE FILTER	18565 / 161561996599	253483 / 304421

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	<0.000001%	0.000000%
2	HUNSPELL ENGLISH LUCENE FILTER	0.583664	0.000023%	83.267252%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	ALL_CANDIDATES	0.999967	1.000000	1.000000	1.000000	1.000000	0.000000
2	HUNSPELL ENGLISH LUCENE FILTER	ALL_CANDIDATES	0.581816	0.167327	1.000000	0.583664	0.999998	0.000002

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	ALL_CANDIDATES	0.999974	0.999984	0.999993	0.999967	0.999984	0.999984
2	HUNSPELL ENGLISH LUCENE FILTER	ALL_CANDIDATES	0.389065	0.259907	0.195130	0.149364	0.312016	0.312015

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	304421	10	0	161561996589	10 / 161561996599	0 / 304421
2	HUNSPELL ENGLISH LUCENE FILTER	ALL_CANDIDATES	50938	36612	253483	161561959987	36612 / 161561996599	253483 / 304421

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	14087	3	7	13355	2.349408%	8	582082
HUNSPELL ENGLISH LUCENE FILTER	5176	1036	17011	5685	1.000104%	4	574142

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: 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
• 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