Ukrainian stemmer benchmarks ★★★★☆

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

Dictionary size: ★★★★☆. The exact count is 14,150 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
uk-ua-default	1.0.0	UK_UA	1,493	14,150	15,737	2,985	12,752	changed tokens	12,752

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 15,737.

Command class	Meaning	Word forms	Share
AppendCharacterCommand	Appends one character to the end of the word form.	267	1.697%
BackwardCompoundCommand	Applies a multi-step backward patch made from skip, delete, insert, and replace operations.	4,156	26.409%
DeleteSuffixCommand	Deletes one or more trailing characters from the word form.	5,883	37.383%
PreserveCommand	Returns the word form unchanged because it already matches the preferred root.	2,962	18.822%
ReplaceLastCharacterCommand	Replaces the final character of the word form.	2,469	15.689%

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	99.307%	99.365%	99.062%	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	86.815%	83.759%	99.866%	Benchmark-only Ukrainian Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene MorfologikFilter	92.362%	90.637%	99.732%	Dictionary-based path; Morfologik can emit multiple terms.
Morfologik direct	92.362%	90.637%	99.732%	Direct dictionary lookup; first returned stem is used for quality when no ranking weight is exposed.

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	ukrainianRadixor	0.625	0.015	49.0	1.000	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	luceneHunspellStemFilter	55.325	7.036	4338.5	88.453	Benchmark-only Ukrainian Hunspell dictionary compared via Lucene HunspellStemFilter.
Morfologik direct	ukrainianMorfologikDirect	8.584	0.122	673.2	13.725	Direct Morfologik dictionary lookup; first returned stem is used for quality.
Lucene MorfologikFilter	ukrainianLuceneMorfologikFilter	15.362	0.300	1204.6	24.560	Dictionary-based Morfologik TokenFilter; may emit multiple terms.

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 uk-ua-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%	14,144	50.982% (50.365–54.332)	42.530% (42.187–47.222)	85.464% (84.667–88.227)
20%	12,528	55.281% (54.207–58.249)	47.743% (46.925–51.631)	87.400% (85.313–88.161)
30%	10,951	58.222% (57.492–59.390)	51.048% (50.564–52.388)	87.380% (86.917–89.403)
40%	9,392	58.933% (57.994–60.117)	51.880% (50.380–53.150)	89.013% (87.486–90.578)
50%	7,838	60.214% (59.232–62.314)	53.268% (51.925–55.648)	90.135% (88.140–91.180)
60%	6,256	61.416% (58.976–62.308)	54.708% (52.041–55.903)	90.211% (87.669–91.525)
70%	4,675	62.139% (60.712–63.062)	55.456% (54.248–56.768)	89.367% (85.811–90.562)
80%	3,166	62.236% (61.727–64.985)	56.335% (54.703–59.359)	88.475% (87.879–93.255)
90%	1,576	63.233% (61.940–65.736)	57.218% (55.192–59.594)	90.476% (88.514–93.243)

Generalization conclusion

• Median exactness on genuinely unseen changed forms moves from 42.530% at 10% training knowledge to 57.218% at 90%, a measured +14.689 percentage-point change.
• Unseen all-form exactness moves from 50.982% at 10% training knowledge to 63.233% at 90%, a measured +12.251 percentage-point change.
• Preservation of unseen already-root forms moves from 85.464% at 10% training knowledge to 90.476% at 90%, a measured +5.012 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 UK_UA 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
1,493	15,737	81.03%	286	7	33.43×	76

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%	43.396%	43.396%	+0.000 pp	0.7798	0.7795	1.000×	0.944×
20%	48.639%	48.639%	+0.000 pp	0.8092	0.8092	1.000×	0.947×
30%	49.864%	49.864%	+0.000 pp	0.8168	0.8168	1.000×	0.942×
40%	51.644%	51.644%	+0.000 pp	0.8298	0.8298	1.000×	0.958×
50%	52.642%	52.642%	+0.000 pp	0.8348	0.8348	1.000×	0.946×
60%	54.724%	54.724%	+0.000 pp	0.8443	0.8443	1.000×	0.934×
70%	56.602%	56.602%	+0.000 pp	0.8487	0.8487	1.000×	0.936×
80%	55.397%	55.397%	+0.000 pp	0.8552	0.8552	1.000×	0.943×
90%	56.474%	56.474%	+0.000 pp	0.8561	0.8561	1.000×	0.949×

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	+1.000	+0.479…+1.000	45
patch_command_ratio	value_references	+1.000	+0.479…+1.000	45
match_cost	patch_command_ratio	-0.217	-0.227…-0.166	45
match_cost	trie_nodes	-0.217	-0.227…-0.166	45
match_cost	value_references	-0.217	-0.227…-0.166	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	17 / 45
average_path_length	unseen_f05	+1.000	-1.000…+1.000	no	19 / 45
trie_edges	unseen_over_percent	-0.641	-0.641…-0.641	no	1 / 45
average_path_length	unseen_under_percent	-1.000	-1.000…+1.000	no	19 / 45

Edit-cost conclusion

• With baseline costs, median unseen changed-form exactness changes from 43.396% at 10% knowledge to 56.474% at 90%, a +13.078 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 5.67% (0.943× baseline).
• Under the selected costs, the 10%–90% knowledge change is +13.078 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 UK_UA 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 uk-ua-default, loaded from classpath resource org/egothor/stemmer/models/uk-ua-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 0.995816 among 4 deterministic stemmers. The runner-up is UKRAINIAN LUCENE MORFOLOGIK FILTER at 0.928906, a difference of 0.066910. This rank does not imply leadership in throughput or every secondary metric.
• LOWERCASE_GROUPS_ONLY: Radixor ranks first by balanced accuracy at 0.995815 among 4 deterministic stemmers. The runner-up is UKRAINIAN LUCENE MORFOLOGIK FILTER at 0.928888, a difference of 0.066926. This rank does not imply leadership in throughput or every secondary metric.

ALL_WORDS

This mode contains 12 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	0.995816	0.000000%	0.836852%
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	0.928906	0.000028%	14.218810%
3	UKRAINIAN MORFOLOGIK DIRECT	0.928783	0.000028%	14.243378%
4	HUNSPELL UKRAINIAN LUCENE FILTER	0.885789	0.000006%	22.842226%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	1.000000	0.991631	1.000000	0.995816	0.999995	0.000005
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	PRIMARY_OUTPUT	0.999499	0.857812	1.000000	0.928906	0.999907	0.000093
3	UKRAINIAN MORFOLOGIK DIRECT	PRIMARY_OUTPUT	0.999499	0.857566	1.000000	0.928783	0.999907	0.000093
4	HUNSPELL UKRAINIAN LUCENE FILTER	PRIMARY_OUTPUT	0.999881	0.771578	1.000000	0.885789	0.999851	0.000149

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.998315	0.995798	0.993294	0.991631	0.995807	0.995804
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	PRIMARY_OUTPUT	0.967537	0.923251	0.882842	0.857443	0.925949	0.925906
3	UKRAINIAN MORFOLOGIK DIRECT	PRIMARY_OUTPUT	0.967474	0.923109	0.882634	0.857198	0.925817	0.925774
4	HUNSPELL UKRAINIAN LUCENE FILTER	PRIMARY_OUTPUT	0.944015	0.871018	0.808499	0.771507	0.878343	0.878277

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	64580	0	545	100039050	0 / 100039050	545 / 65125
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	PRIMARY_OUTPUT	55865	28	9260	100039022	28 / 100039050	9260 / 65125
3	UKRAINIAN MORFOLOGIK DIRECT	PRIMARY_OUTPUT	55849	28	9276	100039022	28 / 100039050	9276 / 65125
4	HUNSPELL UKRAINIAN LUCENE FILTER	PRIMARY_OUTPUT	50249	6	14876	100039044	6 / 100039050	14876 / 65125

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)
HUNSPELL UKRAINIAN LUCENE FILTER	0.000000%	14.533589%
UKRAINIAN LUCENE MORFOLOGIK FILTER	0.000000%	7.594626%
UKRAINIAN MORFOLOGIK DIRECT	0.000000%	7.619194%
Radixor	0.000000%	0.000000%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
HUNSPELL UKRAINIAN LUCENE FILTER	0 / 100039050	9465 / 65125
UKRAINIAN LUCENE MORFOLOGIK FILTER	0 / 100039050	4946 / 65125
UKRAINIAN MORFOLOGIK DIRECT	0 / 100039050	4962 / 65125
Radixor	0 / 100039050	0 / 65125

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	0.000000%	0.000000%
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	0.962027	0.000059%	7.594626%
3	UKRAINIAN MORFOLOGIK DIRECT	0.961904	0.000059%	7.619194%
4	HUNSPELL UKRAINIAN LUCENE FILTER	0.927332	0.000047%	14.533589%

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	UKRAINIAN LUCENE MORFOLOGIK FILTER	ALL_CANDIDATES	0.999021	0.924054	0.999999	0.962027	0.999950	0.000050
3	UKRAINIAN MORFOLOGIK DIRECT	ALL_CANDIDATES	0.999020	0.923808	0.999999	0.961904	0.999950	0.000050
4	HUNSPELL UKRAINIAN LUCENE FILTER	ALL_CANDIDATES	0.999156	0.854664	1.000000	0.927332	0.999905	0.000095

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	UKRAINIAN LUCENE MORFOLOGIK FILTER	ALL_CANDIDATES	0.983070	0.960076	0.938133	0.923217	0.960806	0.960782
3	UKRAINIAN MORFOLOGIK DIRECT	ALL_CANDIDATES	0.983014	0.959943	0.937931	0.922972	0.960678	0.960654
4	HUNSPELL UKRAINIAN LUCENE FILTER	ALL_CANDIDATES	0.966477	0.921279	0.880120	0.854048	0.924090	0.924046

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	65125	0	0	100039050	0 / 100039050	0 / 65125
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	ALL_CANDIDATES	60179	59	4946	100038991	59 / 100039050	4946 / 65125
3	UKRAINIAN MORFOLOGIK DIRECT	ALL_CANDIDATES	60163	59	4962	100038991	59 / 100039050	4962 / 65125
4	HUNSPELL UKRAINIAN LUCENE FILTER	ALL_CANDIDATES	55660	47	9465	100039003	47 / 100039050	9465 / 65125

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
HUNSPELL UKRAINIAN LUCENE FILTER	5411	6	41	1259	8.897527%	6	15577
UKRAINIAN LUCENE MORFOLOGIK FILTER	4314	28	31	2130	15.053004%	6	16748
UKRAINIAN MORFOLOGIK DIRECT	4314	28	31	2130	15.053004%	6	16748
Radixor	545	0	0	95	0.671378%	2	14245

LOWERCASE_GROUPS_ONLY

This mode contains 12 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	0.995815	0.000000%	0.837058%
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	0.928888	0.000028%	14.222304%
3	UKRAINIAN MORFOLOGIK DIRECT	0.928888	0.000028%	14.222304%
4	HUNSPELL UKRAINIAN LUCENE FILTER	0.885791	0.000006%	22.841696%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	1.000000	0.991629	1.000000	0.995815	0.999995	0.000005
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	PRIMARY_OUTPUT	0.999499	0.857777	1.000000	0.928888	0.999907	0.000093
3	UKRAINIAN MORFOLOGIK DIRECT	PRIMARY_OUTPUT	0.999499	0.857777	1.000000	0.928888	0.999907	0.000093
4	HUNSPELL UKRAINIAN LUCENE FILTER	PRIMARY_OUTPUT	0.999881	0.771583	1.000000	0.885791	0.999851	0.000149

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.998315	0.995797	0.993292	0.991629	0.995806	0.995803
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	PRIMARY_OUTPUT	0.967528	0.923231	0.882812	0.857408	0.925930	0.925887
3	UKRAINIAN MORFOLOGIK DIRECT	PRIMARY_OUTPUT	0.967528	0.923231	0.882812	0.857408	0.925930	0.925887
4	HUNSPELL UKRAINIAN LUCENE FILTER	PRIMARY_OUTPUT	0.944017	0.871021	0.808503	0.771512	0.878346	0.878280

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	64564	0	545	99911761	0 / 99911761	545 / 65109
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	PRIMARY_OUTPUT	55849	28	9260	99911733	28 / 99911761	9260 / 65109
3	UKRAINIAN MORFOLOGIK DIRECT	PRIMARY_OUTPUT	55849	28	9260	99911733	28 / 99911761	9260 / 65109
4	HUNSPELL UKRAINIAN LUCENE FILTER	PRIMARY_OUTPUT	50237	6	14872	99911755	6 / 99911761	14872 / 65109

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)
HUNSPELL UKRAINIAN LUCENE FILTER	0.000000%	14.537161%
UKRAINIAN LUCENE MORFOLOGIK FILTER	0.000000%	7.596492%
UKRAINIAN MORFOLOGIK DIRECT	0.000000%	7.596492%
Radixor	0.000000%	0.000000%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
HUNSPELL UKRAINIAN LUCENE FILTER	0 / 99911761	9465 / 65109
UKRAINIAN LUCENE MORFOLOGIK FILTER	0 / 99911761	4946 / 65109
UKRAINIAN MORFOLOGIK DIRECT	0 / 99911761	4946 / 65109
Radixor	0 / 99911761	0 / 65109

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	0.000000%	0.000000%
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	0.962017	0.000059%	7.596492%
3	UKRAINIAN MORFOLOGIK DIRECT	0.962017	0.000059%	7.596492%
4	HUNSPELL UKRAINIAN LUCENE FILTER	0.927314	0.000047%	14.537161%

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	UKRAINIAN LUCENE MORFOLOGIK FILTER	ALL_CANDIDATES	0.999020	0.924035	0.999999	0.962017	0.999950	0.000050
3	UKRAINIAN MORFOLOGIK DIRECT	ALL_CANDIDATES	0.999020	0.924035	0.999999	0.962017	0.999950	0.000050
4	HUNSPELL UKRAINIAN LUCENE FILTER	ALL_CANDIDATES	0.999156	0.854628	1.000000	0.927314	0.999905	0.000095

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	UKRAINIAN LUCENE MORFOLOGIK FILTER	ALL_CANDIDATES	0.983065	0.960066	0.938118	0.923199	0.960796	0.960772
3	UKRAINIAN MORFOLOGIK DIRECT	ALL_CANDIDATES	0.983065	0.960066	0.938118	0.923199	0.960796	0.960772
4	HUNSPELL UKRAINIAN LUCENE FILTER	ALL_CANDIDATES	0.966468	0.921258	0.880089	0.854012	0.924071	0.924027

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	65109	0	0	99911761	0 / 99911761	0 / 65109
2	UKRAINIAN LUCENE MORFOLOGIK FILTER	ALL_CANDIDATES	60163	59	4946	99911702	59 / 99911761	4946 / 65109
3	UKRAINIAN MORFOLOGIK DIRECT	ALL_CANDIDATES	60163	59	4946	99911702	59 / 99911761	4946 / 65109
4	HUNSPELL UKRAINIAN LUCENE FILTER	ALL_CANDIDATES	55644	47	9465	99911714	47 / 99911761	9465 / 65109

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
HUNSPELL UKRAINIAN LUCENE FILTER	5407	6	41	1258	8.896118%	6	15567
UKRAINIAN LUCENE MORFOLOGIK FILTER	4314	28	31	2130	15.062584%	6	16739
UKRAINIAN MORFOLOGIK DIRECT	4314	28	31	2130	15.062584%	6	16739
Radixor	545	0	0	95	0.671805%	2	14236

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: UK_UA
• 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