Czech stemmer benchmarks ★★★★☆

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

Dictionary size: ★★★★☆. The exact count is 51,401 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
cs-cz-default	1.0.0	CS_CZ	5,113	51,401	56,612	10,049	46,563	changed tokens	46,563

Radixor patch-command distribution

Radixor stores the preferred transformation for each normalized dictionary word form as a compiled patch command. This distribution shows which runtime command class is selected by the trained trie for the complete default-model dictionary. The total number of preferred patch commands analyzed for this language is 56,612.

Command class	Meaning	Word forms	Share
AppendCharacterCommand	Appends one character to the end of the word form.	711	1.256%
BackwardCompoundCommand	Applies a multi-step backward patch made from skip, delete, insert, and replace operations.	22,643	39.997%
DeleteSuffixCommand	Deletes one or more trailing characters from the word form.	15,007	26.509%
PreserveCommand	Returns the word form unchanged because it already matches the preferred root.	10,046	17.745%
ReplaceLastCharacterCommand	Replaces the final character of the word form.	8,205	14.493%

Accuracy

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	99.465%	99.439%	99.582%	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	84.850%	82.269%	96.806%	Benchmark-only Czech Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene CzechStemFilter	16.784%	15.538%	22.559%	Lucene Czech suffix stemmer implemented as a TokenFilter.
Official Snowball direct (Java)	19.865%	18.186%	27.645%	Official Snowball 3.1.0 generated Java stemmer; rule-based suffix algorithm.

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[cs-cz-default]	4.255	0.770	91.4	1.000	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	luceneHunspellStemFilter	428.813	60.284	9209.3	100.777	Benchmark-only Czech Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene CzechStemFilter	czechLuceneCzechStemFilter	3.046	0.076	65.4	0.716	Czech suffix stemmer implemented as a Lucene TokenFilter.
Official Snowball direct (Java)	snowballDirect[CZECH]	4.620	0.737	99.2	1.086	Official Snowball 3.1.0 generated Java stemmer; direct API. The reported Radixor and Snowball 99.9% JMH intervals overlap; this summary does not support a directional conclusion, and the runtime ratio remains a point estimate.

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 cs-cz-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%	50,905	69.921% (66.235–70.712)	65.439% (60.991–66.671)	89.636% (88.686–90.665)
20%	45,111	72.003% (69.047–72.412)	68.155% (64.204–68.726)	89.767% (89.232–91.428)
30%	39,562	73.325% (71.387–74.104)	69.913% (67.160–70.569)	90.571% (89.035–90.982)
40%	33,944	74.037% (73.092–74.837)	70.380% (69.385–71.348)	90.721% (89.903–91.126)
50%	28,195	74.013% (73.648–74.299)	70.612% (69.908–70.743)	91.128% (89.182–91.481)
60%	22,567	74.064% (73.458–75.665)	70.641% (69.532–72.243)	91.201% (89.195–91.780)
70%	16,910	74.517% (73.186–75.890)	70.679% (69.464–72.757)	91.409% (90.430–92.421)
80%	11,295	74.375% (72.961–75.341)	70.977% (68.757–71.820)	91.822% (89.763–92.700)
90%	5,497	73.270% (71.970–75.506)	69.900% (67.596–72.033)	91.296% (89.606–92.449)

Generalization conclusion

• Median exactness on genuinely unseen changed forms moves from 65.439% at 10% training knowledge to 69.900% at 90%, a measured +4.461 percentage-point change.
• Unseen all-form exactness moves from 69.921% at 10% training knowledge to 73.270% at 90%, a measured +3.349 percentage-point change.
• Preservation of unseen already-root forms moves from 89.636% at 10% training knowledge to 91.296% at 90%, a measured +1.660 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 CS_CZ 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
5,113	56,612	82.25%	537	12	19.50×	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%	64.331%	64.331%	+0.000 pp	0.9049	0.9049	1.000×	0.953×
20%	66.560%	66.560%	+0.000 pp	0.9085	0.9085	1.000×	0.935×
30%	68.753%	68.753%	+0.000 pp	0.9203	0.9203	1.000×	0.915×
40%	69.876%	69.876%	+0.000 pp	0.9233	0.9233	1.000×	0.928×
50%	70.887%	70.887%	+0.000 pp	0.9279	0.9278	1.000×	0.931×
60%	70.925%	70.925%	+0.000 pp	0.9300	0.9300	1.000×	0.940×
70%	71.428%	71.428%	+0.000 pp	0.9327	0.9327	1.000×	0.931×
80%	71.146%	71.146%	+0.000 pp	0.9310	0.9310	1.000×	0.952×
90%	70.970%	70.970%	+0.000 pp	0.9300	0.9300	1.000×	0.957×

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	trie_nodes	+0.932	+0.478…+1.000	45
patch_command_ratio	value_references	+0.965	+0.477…+1.000	45
replace_cost	trie_edges	+0.741	+0.135…+0.775	45
match_cost	trie_edges	-0.145	-0.189…-0.132	45
match_cost	dense_table_slots	-0.146	-0.205…-0.132	45
match_cost	patch_command_ratio	-0.151	-0.217…-0.130	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	-1.000	-1.000…+0.875	no	41 / 45
replace_to_delete_insert	unseen_over_percent	-0.730	-0.876…-0.396	no	17 / 45
average_path_length	unseen_under_percent	+1.000	-1.000…+1.000	no	41 / 45

Edit-cost conclusion

• With baseline costs, median unseen changed-form exactness changes from 64.331% at 10% knowledge to 70.970% at 90%, a +6.639 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 5.83% (0.942× baseline).
• Under the selected costs, the 10%–90% knowledge change is +6.639 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 CS_CZ 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 cs-cz-default, loaded from classpath resource org/egothor/stemmer/models/cs-cz-default/stemmer.gz. The following findings compare only deterministic PRIMARY_OUTPUT rows over identical included groups; candidate policies are reported separately as capability analyses.

• ALL_WORDS: Radixor ranks first by balanced accuracy at 0.996617 among 4 deterministic stemmers. The runner-up is HUNSPELL CZECH LUCENE FILTER at 0.854132, a difference of 0.142485. This rank does not imply leadership in throughput or every secondary metric.
• LOWERCASE_GROUPS_ONLY: Radixor ranks first by balanced accuracy at 0.997195 among 4 deterministic stemmers. The runner-up is HUNSPELL CZECH LUCENE FILTER at 0.853150, a difference of 0.144045. This rank does not imply leadership in throughput or every secondary metric.

ALL_WORDS

This mode contains 8 result rows, 4 evaluated stemmers, and 3 output policies. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. 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	0.996617	0.000000%	0.676519%
2	HUNSPELL CZECH LUCENE FILTER	0.854132	0.000691%	29.172837%
3	CZECH LUCENE CZECH STEM FILTER	0.794343	0.000928%	41.130549%
4	SNOWBALL CZECH DIRECT	0.786366	0.000904%	42.725842%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	1.000000	0.993235	1.000000	0.996617	0.999998	0.000002
2	HUNSPELL CZECH LUCENE FILTER	PRIMARY_OUTPUT	0.958877	0.708272	0.999993	0.854132	0.999927	0.000073
3	CZECH LUCENE CZECH STEM FILTER	PRIMARY_OUTPUT	0.935210	0.588695	0.999991	0.794343	0.999897	0.000103
4	SNOWBALL CZECH DIRECT	PRIMARY_OUTPUT	0.935153	0.572742	0.999991	0.786366	0.999894	0.000106

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.998640	0.996606	0.994581	0.993235	0.996612	0.996611
2	HUNSPELL CZECH LUCENE FILTER	PRIMARY_OUTPUT	0.895506	0.814739	0.747335	0.687392	0.824103	0.824070
3	CZECH LUCENE CZECH STEM FILTER	PRIMARY_OUTPUT	0.836710	0.722556	0.635811	0.565626	0.741992	0.741949
4	SNOWBALL CZECH DIRECT	PRIMARY_OUTPUT	0.830101	0.710396	0.620864	0.550864	0.731848	0.731804

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	298476	0	2033	1320705191	0 / 1320705191	2033 / 300509
2	HUNSPELL CZECH LUCENE FILTER	PRIMARY_OUTPUT	212842	9128	87667	1320696063	9128 / 1320705191	87667 / 300509
3	CZECH LUCENE CZECH STEM FILTER	PRIMARY_OUTPUT	176908	12256	123601	1320692935	12256 / 1320705191	123601 / 300509
4	SNOWBALL CZECH DIRECT	PRIMARY_OUTPUT	172114	11935	128395	1320693256	11935 / 1320705191	128395 / 300509

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 CZECH LUCENE FILTER	0.000650%	25.611213%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 1320705191	0 / 300509
HUNSPELL CZECH LUCENE FILTER	8582 / 1320705191	76964 / 300509

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	0.000000%	0.000000%
2	HUNSPELL CZECH LUCENE FILTER	0.871940	0.000816%	25.611213%

Classification metrics

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
2	HUNSPELL CZECH LUCENE FILTER	ALL_CANDIDATES	0.954016	0.743888	0.999992	0.871940	0.999934	0.000066

Pair-relation metrics

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
2	HUNSPELL CZECH LUCENE FILTER	ALL_CANDIDATES	0.903001	0.835949	0.778167	0.718138	0.842426	0.842395

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	300509	0	0	1320705191	0 / 1320705191	0 / 300509
2	HUNSPELL CZECH LUCENE FILTER	ALL_CANDIDATES	223545	10775	76964	1320694416	10775 / 1320705191	76964 / 300509

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	2033	0	0	321	0.624501%	4	51739
HUNSPELL CZECH LUCENE FILTER	10703	546	1647	3194	6.213887%	5	55179

LOWERCASE_GROUPS_ONLY

This mode contains 8 result rows, 4 evaluated stemmers, and 3 output policies. Applied-row and form counts are shown per row because adapters share the language corpus but policy rows remain independently auditable. 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	0.997195	0.000000%	0.561033%
2	HUNSPELL CZECH LUCENE FILTER	0.853150	0.000700%	29.369351%
3	CZECH LUCENE CZECH STEM FILTER	0.792522	0.000918%	41.494586%
4	SNOWBALL CZECH DIRECT	0.784821	0.000923%	43.034822%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	1.000000	0.994390	1.000000	0.997195	0.999999	0.000001
2	HUNSPELL CZECH LUCENE FILTER	PRIMARY_OUTPUT	0.958957	0.706306	0.999993	0.853150	0.999925	0.000075
3	CZECH LUCENE CZECH STEM FILTER	PRIMARY_OUTPUT	0.936557	0.585054	0.999991	0.792522	0.999895	0.000105
4	SNOWBALL CZECH DIRECT	PRIMARY_OUTPUT	0.934577	0.569652	0.999991	0.784821	0.999891	0.000109

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.998873	0.997187	0.995507	0.994390	0.997191	0.997190
2	HUNSPELL CZECH LUCENE FILTER	PRIMARY_OUTPUT	0.894932	0.813466	0.745594	0.685581	0.822993	0.822960
3	CZECH LUCENE CZECH STEM FILTER	PRIMARY_OUTPUT	0.836092	0.720206	0.632534	0.562751	0.740227	0.740184
4	SNOWBALL CZECH DIRECT	PRIMARY_OUTPUT	0.828436	0.707849	0.617907	0.547807	0.729646	0.729601

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	295818	0	1669	1284770069	0 / 1284770069	1669 / 297487
2	HUNSPELL CZECH LUCENE FILTER	PRIMARY_OUTPUT	210117	8993	87370	1284761076	8993 / 1284770069	87370 / 297487
3	CZECH LUCENE CZECH STEM FILTER	PRIMARY_OUTPUT	174046	11790	123441	1284758279	11790 / 1284770069	123441 / 297487
4	SNOWBALL CZECH DIRECT	PRIMARY_OUTPUT	169464	11863	128023	1284758206	11863 / 1284770069	128023 / 297487

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 CZECH LUCENE FILTER	0.000663%	25.840457%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 1284770069	0 / 297487
HUNSPELL CZECH LUCENE FILTER	8518 / 1284770069	76872 / 297487

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	0.000000%	0.000000%
2	HUNSPELL CZECH LUCENE FILTER	0.870794	0.000819%	25.840457%

Classification metrics

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
2	HUNSPELL CZECH LUCENE FILTER	ALL_CANDIDATES	0.954473	0.741595	0.999992	0.870794	0.999932	0.000068

Pair-relation metrics

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
2	HUNSPELL CZECH LUCENE FILTER	ALL_CANDIDATES	0.902651	0.834675	0.776220	0.716259	0.841328	0.841297

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	297487	0	0	1284770069	0 / 1284770069	0 / 297487
2	HUNSPELL CZECH LUCENE FILTER	ALL_CANDIDATES	220615	10523	76872	1284759546	10523 / 1284770069	76872 / 297487

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	1669	0	0	269	0.530603%	4	50975
HUNSPELL CZECH LUCENE FILTER	10498	475	1530	3117	6.148293%	5	54394

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: CS_CZ
• Processing modes: ALL_WORDS, LOWERCASE_GROUPS_ONLY
• Stemmer versions and transitive artifacts: resolved by the repository's JMH Gradle configuration and gradle.lockfile
• 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