French stemmer benchmarks ★★★★★

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

Dictionary size: ★★★★★. The exact count is 404,011 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
fr-fr-default	1.0.0	FR_FR	59,240	404,011	474,110	108,141	365,969	changed tokens	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 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.	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

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	94.831%	94.859%	94.734%	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	68.923%	63.617%	86.876%	Benchmark-only French Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene FrenchMinimalStemFilter	11.472%	6.236%	29.192%	Minimal suffix reducer; narrow baseline, not a full stemmer.
Lucene SnowballFilter	8.551%	5.183%	19.952%	Lucene TokenFilter integration path around the Snowball algorithm.
Official Snowball direct (Java)	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 Radixor's dictionary-trained transformation model.

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.

Stemmer	Benchmark method	Score ms/op	Error ms	ns/token	Relative vs Radixor	Note
Radixor	radixor[fr-fr-default]	59.309	1.798	162.1	1.000	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	luceneHunspellStemFilter	1763.422	98.692	4818.5	29.733	Benchmark-only French Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene FrenchMinimalStemFilter	frenchLuceneFrenchMinimalStemFilter	18.597	0.435	50.8	0.314	Minimal French suffix reducer; narrow baseline.
Lucene FrenchLightStemFilter	frenchLuceneFrenchLightStemFilter	27.816	0.491	76.0	0.469	Light French suffix stemmer.
Official Snowball direct (Java)	snowballDirect[FRENCH]	117.446	6.404	320.9	1.980	Official Snowball generated Java stemmer; direct API.
Lucene SnowballFilter	luceneSnowballFilter[FRENCH]	132.498	6.874	362.0	2.234	Lucene TokenFilter path around Snowball; includes TokenStream overhead.

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 fr-fr-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%	420,578	74.276% (73.276–74.926)	69.627% (68.400–70.528)	90.013% (89.987–90.359)
20%	369,337	76.769% (76.583–76.946)	72.504% (72.339–72.772)	91.463% (90.996–91.884)
30%	318,984	78.161% (78.141–78.274)	74.079% (73.976–74.276)	92.537% (92.395–92.906)
40%	270,536	79.255% (79.039–79.552)	75.314% (75.088–75.702)	93.386% (93.086–93.418)
50%	222,533	80.344% (79.709–80.528)	76.569% (75.778–76.888)	94.006% (93.785–94.099)
60%	175,326	81.224% (80.630–81.431)	77.560% (76.779–77.867)	94.813% (94.641–94.831)
70%	130,120	82.101% (81.643–82.329)	78.644% (78.015–78.839)	95.431% (95.257–95.560)
80%	85,315	82.721% (81.861–83.091)	79.345% (78.141–79.693)	96.074% (95.896–96.266)
90%	41,750	83.469% (82.278–83.711)	80.031% (78.595–80.412)	96.664% (96.351–96.813)

Generalization conclusion

• Median exactness on genuinely unseen changed forms moves from 69.627% at 10% training knowledge to 80.031% at 90%, a measured +10.404 percentage-point change.
• Unseen all-form exactness moves from 74.276% at 10% training knowledge to 83.469% at 90%, a measured +9.193 percentage-point change.
• Preservation of unseen already-root forms moves from 90.013% at 10% training knowledge to 96.664% at 90%, a measured +6.651 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 FR_FR 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
59,240	474,110	77.19%	2,730	32	7.31×	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%	69.937%	69.937%	+0.000 pp	0.8816	0.8816	1.000×	0.718×
20%	72.375%	72.375%	+0.000 pp	0.8939	0.8939	1.000×	0.721×
30%	74.543%	74.542%	-0.000 pp	0.9055	0.9055	1.000×	0.720×
40%	75.720%	75.720%	+0.000 pp	0.9100	0.9100	1.000×	0.724×
50%	76.846%	76.846%	+0.000 pp	0.9156	0.9156	1.000×	0.732×
60%	77.664%	77.664%	+0.000 pp	0.9201	0.9201	1.000×	0.734×
70%	78.435%	78.435%	+0.000 pp	0.9239	0.9239	1.000×	0.744×
80%	79.746%	79.746%	+0.000 pp	0.9289	0.9289	1.000×	0.741×
90%	80.058%	80.058%	+0.000 pp	0.9320	0.9320	1.000×	0.750×

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
replace_to_delete_insert	patch_command_ratio	+0.956	+0.943…+0.959	45
patch_command_ratio	value_references	+0.972	+0.934…+1.000	45
patch_command_ratio	trie_nodes	+0.948	+0.931…+0.999	45
replace_to_delete_insert	value_references	+0.924	+0.882…+0.957	45
replace_to_delete_insert	trie_nodes	+0.899	+0.882…+0.956	45
patch_command_ratio	trie_edges	+0.821	+0.814…+0.913	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
average_path_length	unseen_changed_exact	+1.000	-1.000…+1.000	no	38 / 45
average_path_length	unseen_f05	+0.822	-0.902…+1.000	no	39 / 45
value_references	unseen_over_percent	-0.846	-0.984…-0.723	no	18 / 45
logical_leaf_paths	unseen_under_percent	+0.853	+0.692…+1.000	no	39 / 45

Edit-cost conclusion

• With baseline costs, median unseen changed-form exactness changes from 69.937% at 10% knowledge to 80.058% at 90%, a +10.121 pp measured knowledge effect.
• The predeclared selection is D10I10R1M1. Its median unseen changed-form exactness differs from baseline by +0.000 pp and it reduces the median retained-command count by 26.64% (0.734× baseline).
• Under the selected costs, the 10%–90% knowledge change is +10.121 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.
• 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 FR_FR 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 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 0.958627 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. 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	0.958627	<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%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
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

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
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

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
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

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%	0.004320%
HUNSPELL FRENCH LUCENE FILTER	0.000539%	33.189869%

Oracle-bound pair counts

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

ALL_CANDIDATES ranking

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%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
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

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
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

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
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

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	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. 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	0.958856	<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%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
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

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
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

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
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

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 FRENCH LUCENE FILTER	0.000539%	33.211718%

Oracle-bound pair counts

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

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	<0.000001%	0.000000%
2	HUNSPELL FRENCH LUCENE FILTER	0.833938	0.000614%	33.211718%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
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

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
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

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
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

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	440750	1	74	20611	5.143594%	56	422336
HUNSPELL FRENCH LUCENE FILTER	184645	1047	59168	8194	2.044860%	4	409028

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: 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
• 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