German stemmer benchmarks ★★★★★

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

Dictionary size: ★★★★★. The exact count is 277,266 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
de-de-default	1.0.0	DE_DE	54,092	277,266	333,036	90,535	242,501	changed tokens	242,501

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 333,036.

Command class	Meaning	Word forms	Share
AppendCharacterCommand	Appends one character to the end of the word form.	12,107	3.635%
BackwardCompoundCommand	Applies a multi-step backward patch made from skip, delete, insert, and replace operations.	81,805	24.563%
DeleteSuffixCommand	Deletes one or more trailing characters from the word form.	142,376	42.751%
PreserveCommand	Returns the word form unchanged because it already matches the preferred root.	88,820	26.670%
ReplaceLastCharacterCommand	Replaces the final character of the word form.	7,928	2.381%

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	92.725%	92.847%	92.396%	Radixor dictionary-trained patch-command stemmer.
Lucene HunspellStemFilter	47.064%	29.661%	93.678%	Benchmark-only German Hunspell dictionary compared via Lucene HunspellStemFilter.
CISTEM (German)	24.675%	23.724%	27.222%	Benchmark-only CISTEM implementation.
Lucene GermanLightStemFilter	37.434%	35.465%	42.707%	Light suffix stemmer; intentionally narrower than Radixor's lexicon-trained transformation model.
Lucene GermanMinimalStemFilter	27.640%	24.951%	34.844%	Minimal suffix reducer; narrow baseline, not a full stemmer.
Lucene SnowballFilter	30.956%	28.853%	36.589%	Lucene TokenFilter integration path around the Snowball algorithm.
Official Snowball direct (Java)	30.483%	29.030%	34.376%	Official Snowball generated Java stemmer; rule-based suffix algorithm.
Lucene GermanStemFilter	21.559%	19.312%	27.576%	German Lucene stemming TokenFilter; broader than minimal/light variants.

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[de-de-default]	45.882	3.967	189.2	1.000	Radixor dictionary-trained patch-command stemmer.
CISTEM	germanCistem	296.652	6.437	1223.3	6.465	Benchmark-only CISTEM implementation.
Lucene HunspellStemFilter	luceneHunspellStemFilter	315.302	48.761	1300.2	6.872	Benchmark-only German Hunspell dictionary compared via Lucene HunspellStemFilter.
Lucene GermanMinimalStemFilter	germanLuceneGermanMinimalStemFilter	23.069	0.390	95.1	0.503	Minimal German suffix reduction; narrow baseline.
Lucene GermanLightStemFilter	germanLuceneGermanLightStemFilter	23.770	0.569	98.0	0.518	Light German suffix stemmer; narrower than Radixor's lexicon-trained transformation model.
Lucene GermanStemFilter	germanLuceneGermanStemFilter	67.521	0.982	278.4	1.472	Older German stemming TokenFilter with normalization requirements.
Lucene SnowballFilter	luceneSnowballFilter[GERMAN]	105.967	4.323	437.0	2.310	Lucene TokenFilter path around Snowball; includes TokenStream overhead.
Official Snowball direct (Java)	snowballDirect[GERMAN]	99.438	3.835	410.1	2.167	Official Snowball generated Java stemmer; direct API.

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 de-de-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%	293,618	59.280% (58.219–59.702)	49.432% (48.026–49.731)	86.150% (85.836–86.733)
20%	255,708	61.569% (60.724–61.785)	51.959% (50.533–52.161)	88.244% (88.017–88.723)
30%	219,296	62.696% (62.282–63.101)	53.083% (52.613–53.663)	89.436% (88.996–89.500)
40%	184,273	63.564% (63.219–63.981)	54.206% (53.646–54.739)	90.246% (89.885–90.446)
50%	151,254	64.493% (64.083–64.603)	55.371% (54.795–55.510)	90.896% (90.661–91.008)
60%	118,916	65.062% (64.946–65.321)	56.115% (55.912–56.303)	91.372% (91.222–91.619)
70%	87,177	65.784% (65.442–66.138)	56.948% (56.648–57.396)	91.897% (91.743–91.984)
80%	57,109	66.531% (65.981–66.706)	58.016% (57.177–58.316)	92.441% (91.771–92.581)
90%	27,967	67.322% (66.336–67.716)	59.108% (57.552–59.642)	93.117% (92.600–93.345)

Generalization conclusion

• Median exactness on genuinely unseen changed forms moves from 49.432% at 10% training knowledge to 59.108% at 90%, a measured +9.676 percentage-point change.
• Unseen all-form exactness moves from 59.280% at 10% training knowledge to 67.322% at 90%, a measured +8.042 percentage-point change.
• Preservation of unseen already-root forms moves from 86.150% at 10% training knowledge to 93.117% at 90%, a measured +6.967 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 DE_DE 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
54,092	333,036	72.82%	6,986	44	5.32×	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%	48.338%	48.338%	+0.000 pp	0.7731	0.7730	1.000×	0.971×
20%	51.523%	51.523%	+0.001 pp	0.7949	0.7948	1.000×	0.970×
30%	53.276%	53.276%	+0.000 pp	0.8059	0.8059	1.000×	0.969×
40%	53.824%	53.824%	+0.000 pp	0.8136	0.8137	1.000×	0.969×
50%	55.399%	55.399%	+0.000 pp	0.8263	0.8263	1.000×	0.970×
60%	56.171%	56.171%	+0.000 pp	0.8337	0.8338	1.000×	0.971×
70%	56.757%	56.757%	+0.000 pp	0.8373	0.8374	1.000×	0.971×
80%	57.857%	57.857%	+0.000 pp	0.8456	0.8456	1.000×	0.971×
90%	58.463%	58.463%	+0.000 pp	0.8496	0.8496	1.000×	0.972×

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	+0.995	+0.993…+1.000	45
patch_command_ratio	trie_nodes	+0.972	+0.949…+0.998	45
replace_to_delete_insert	dense_table_slots	+0.687	+0.650…+0.852	45
replace_to_delete_insert	trie_edges	+0.666	+0.602…+0.711	45
replace_cost	dense_table_slots	+0.626	+0.601…+0.744	45
replace_cost	trie_edges	+0.596	+0.565…+0.629	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
logical_leaf_paths	unseen_changed_exact	-0.748	-0.895…-0.227	no	39 / 45
trie_nodes	unseen_f05	-0.611	-0.964…+0.063	no	45 / 45
replace_to_delete_insert	unseen_over_percent	-0.730	-0.954…+0.737	no	37 / 45
logical_leaf_paths	unseen_under_percent	+0.573	-0.494…+0.899	no	45 / 45

Edit-cost conclusion

• With baseline costs, median unseen changed-form exactness changes from 48.338% at 10% knowledge to 58.463% at 90%, a +10.125 pp measured knowledge effect.
• The predeclared selection is D2I5R3M0. Its median unseen changed-form exactness differs from baseline by +0.000 pp and it reduces the median retained-command count by 2.94% (0.971× baseline).
• Under the selected costs, the 10%–90% knowledge change is +10.125 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 DE_DE 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 de-de-default, loaded from classpath resource org/egothor/stemmer/models/de-de-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 0.910445 among 8 deterministic stemmers. The runner-up is GERMAN CISTEM at 0.878527, a difference of 0.031918. This rank does not imply leadership in throughput or every secondary metric.
• LOWERCASE_GROUPS_ONLY: Radixor ranks first by balanced accuracy at 0.966959 among 8 deterministic stemmers. The runner-up is GERMAN CISTEM at 0.914727, a difference of 0.052232. This rank does not imply leadership in throughput or every secondary metric.

ALL_WORDS

This mode contains 12 result rows, 8 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	0.910445	0.000002%	17.910967%
2	GERMAN CISTEM	0.878527	0.000674%	24.293900%
3	SNOWBALL GERMAN DIRECT	0.776012	0.000171%	44.797420%
4	SNOWBALL GERMAN LUCENE FILTER	0.769071	0.000371%	46.185528%
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	0.753833	0.000191%	49.233299%
6	GERMAN LUCENE GERMAN STEM FILTER	0.720992	0.000443%	55.801084%
7	HUNSPELL GERMAN LUCENE FILTER	0.640308	0.000290%	71.938102%
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	0.595748	0.000088%	80.850384%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	0.999400	0.820890	1.000000	0.910445	0.999994	0.000006
2	GERMAN CISTEM	PRIMARY_OUTPUT	0.797231	0.757061	0.999993	0.878527	0.999985	0.000015
3	SNOWBALL GERMAN DIRECT	PRIMARY_OUTPUT	0.918571	0.552026	0.999998	0.776012	0.999983	0.000017
4	SNOWBALL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.835220	0.538145	0.999996	0.769071	0.999980	0.000020
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	PRIMARY_OUTPUT	0.902792	0.507667	0.999998	0.753833	0.999981	0.000019
6	GERMAN LUCENE GERMAN STEM FILTER	PRIMARY_OUTPUT	0.777304	0.441989	0.999996	0.720992	0.999976	0.000024
7	HUNSPELL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.771720	0.280619	0.999997	0.640308	0.999972	0.000028
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	PRIMARY_OUTPUT	0.883845	0.191496	0.999999	0.595748	0.999971	0.000029

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.957746	0.901392	0.851302	0.820486	0.905758	0.905755
2	GERMAN CISTEM	PRIMARY_OUTPUT	0.788860	0.776627	0.764768	0.634824	0.776886	0.776879
3	SNOWBALL GERMAN DIRECT	PRIMARY_OUTPUT	0.810886	0.689618	0.599903	0.526272	0.712092	0.712085
4	SNOWBALL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.752175	0.654552	0.579359	0.486494	0.670425	0.670416
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	PRIMARY_OUTPUT	0.781189	0.649884	0.556368	0.481355	0.676991	0.676984
6	GERMAN LUCENE GERMAN STEM FILTER	PRIMARY_OUTPUT	0.674901	0.563540	0.483723	0.392311	0.586140	0.586130
7	HUNSPELL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.571639	0.411577	0.321543	0.259110	0.465359	0.465349
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	PRIMARY_OUTPUT	0.512941	0.314789	0.227071	0.186795	0.411404	0.411396

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	1103976	663	240876	38436733230	663 / 38436733893	240876 / 1344852
2	GERMAN CISTEM	PRIMARY_OUTPUT	1018135	258954	326717	38436474939	258954 / 38436733893	326717 / 1344852
3	SNOWBALL GERMAN DIRECT	PRIMARY_OUTPUT	742393	65811	602459	38436668082	65811 / 38436733893	602459 / 1344852
4	SNOWBALL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	723725	142783	621127	38436591110	142783 / 38436733893	621127 / 1344852
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	PRIMARY_OUTPUT	682737	73514	662115	38436660379	73514 / 38436733893	662115 / 1344852
6	GERMAN LUCENE GERMAN STEM FILTER	PRIMARY_OUTPUT	594410	170297	750442	38436563596	170297 / 38436733893	750442 / 1344852
7	HUNSPELL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	377391	111635	967461	38436622258	111635 / 38436733893	967461 / 1344852
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	PRIMARY_OUTPUT	257534	33845	1087318	38436700048	33845 / 38436733893	1087318 / 1344852

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	0.000001%	8.261653%
HUNSPELL GERMAN LUCENE FILTER	0.000216%	70.811435%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	502 / 38436733893	111107 / 1344852
HUNSPELL GERMAN LUCENE FILTER	83073 / 38436733893	952309 / 1344852

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	0.958692	0.000018%	8.261653%
2	HUNSPELL GERMAN LUCENE FILTER	0.645941	0.000354%	70.811435%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	ALL_CANDIDATES	0.994469	0.917383	1.000000	0.958692	0.999997	0.000003
2	HUNSPELL GERMAN LUCENE FILTER	ALL_CANDIDATES	0.742744	0.291886	0.999996	0.645941	0.999972	0.000028

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	ALL_CANDIDATES	0.978033	0.954372	0.931829	0.912726	0.955149	0.955147
2	HUNSPELL GERMAN LUCENE FILTER	ALL_CANDIDATES	0.567444	0.419080	0.332218	0.265086	0.465614	0.465603

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	1233745	6862	111107	38436727031	6862 / 38436733893	111107 / 1344852
2	HUNSPELL GERMAN LUCENE FILTER	ALL_CANDIDATES	392543	135961	952309	38436597932	135961 / 38436733893	952309 / 1344852

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	129769	161	6199	29035	10.471893%	8	313927
HUNSPELL GERMAN LUCENE FILTER	15152	28562	24326	6482	2.337827%	3	283881

LOWERCASE_GROUPS_ONLY

This mode contains 12 result rows, 8 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	0.966959	0.000001%	6.608210%
2	GERMAN CISTEM	0.914727	0.000812%	17.053716%
3	SNOWBALL GERMAN DIRECT	0.794997	0.000391%	41.000236%
4	SNOWBALL GERMAN LUCENE FILTER	0.774716	0.000325%	45.056540%
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	0.768968	0.000130%	46.206331%
6	GERMAN LUCENE GERMAN STEM FILTER	0.716147	0.000358%	56.770194%
7	HUNSPELL GERMAN LUCENE FILTER	0.659574	0.000556%	68.084626%
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	0.574999	0.000045%	85.000064%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	PRIMARY_OUTPUT	0.999900	0.933918	1.000000	0.966959	0.999995	0.000005
2	GERMAN CISTEM	PRIMARY_OUTPUT	0.892172	0.829463	0.999992	0.914727	0.999978	0.000022
3	SNOWBALL GERMAN DIRECT	PRIMARY_OUTPUT	0.924304	0.589998	0.999996	0.794997	0.999963	0.000037
4	SNOWBALL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.931871	0.549435	0.999997	0.774716	0.999960	0.000040
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	PRIMARY_OUTPUT	0.971001	0.537937	0.999999	0.768968	0.999961	0.000039
6	GERMAN LUCENE GERMAN STEM FILTER	PRIMARY_OUTPUT	0.907196	0.432298	0.999996	0.716147	0.999950	0.000050
7	HUNSPELL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.823043	0.319154	0.999994	0.659574	0.999939	0.000061
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	PRIMARY_OUTPUT	0.964480	0.149999	1.000000	0.574999	0.999931	0.000069

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	PRIMARY_OUTPUT	0.985968	0.965783	0.946408	0.933831	0.966346	0.966343
2	GERMAN CISTEM	PRIMARY_OUTPUT	0.878883	0.859676	0.841289	0.753887	0.860246	0.860236
3	SNOWBALL GERMAN DIRECT	PRIMARY_OUTPUT	0.830220	0.720249	0.636004	0.562804	0.738469	0.738454
4	SNOWBALL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.817997	0.691285	0.598564	0.528217	0.715543	0.715527
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	PRIMARY_OUTPUT	0.836342	0.692324	0.590620	0.529431	0.722729	0.722714
6	GERMAN LUCENE GERMAN STEM FILTER	PRIMARY_OUTPUT	0.743781	0.585563	0.482850	0.413990	0.626242	0.626223
7	HUNSPELL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	0.625524	0.459951	0.363685	0.298660	0.512520	0.512499
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	PRIMARY_OUTPUT	0.462363	0.259621	0.180482	0.149175	0.380357	0.380343

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	PRIMARY_OUTPUT	801691	80	56726	10594963454	80 / 10594963534	56726 / 858417
2	GERMAN CISTEM	PRIMARY_OUTPUT	712025	86055	146392	10594877479	86055 / 10594963534	146392 / 858417
3	SNOWBALL GERMAN DIRECT	PRIMARY_OUTPUT	506464	41477	351953	10594922057	41477 / 10594963534	351953 / 858417
4	SNOWBALL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	471644	34482	386773	10594929052	34482 / 10594963534	386773 / 858417
5	GERMAN LUCENE GERMAN LIGHT STEM FILTER	PRIMARY_OUTPUT	461774	13791	396643	10594949743	13791 / 10594963534	396643 / 858417
6	GERMAN LUCENE GERMAN STEM FILTER	PRIMARY_OUTPUT	371092	37962	487325	10594925572	37962 / 10594963534	487325 / 858417
7	HUNSPELL GERMAN LUCENE FILTER	PRIMARY_OUTPUT	273967	58904	584450	10594904630	58904 / 10594963534	584450 / 858417
8	GERMAN LUCENE GERMAN MINIMAL STEM FILTER	PRIMARY_OUTPUT	128762	4742	729655	10594958792	4742 / 10594963534	729655 / 858417

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 GERMAN LUCENE FILTER	0.000383%	66.866802%

Oracle-bound pair counts

Stemmer	Unavoidable over errors / gold-negative pairs	Unrepairable under errors / gold-related pairs
Radixor	0 / 10594963534	0 / 858417
HUNSPELL GERMAN LUCENE FILTER	40608 / 10594963534	573996 / 858417

ALL_CANDIDATES ranking

Rank	Stemmer	Balanced accuracy	Over-stemming (OI)	Under-stemming (UI)
1	Radixor	1.000000	0.000014%	0.000000%
2	HUNSPELL GERMAN LUCENE FILTER	0.665663	0.000629%	66.866802%

Classification metrics

Rank	Stemmer	Output policy	Precision	Recall	Specificity	Balanced accuracy	Pairwise accuracy	Error rate
1	Radixor	ALL_CANDIDATES	0.998267	1.000000	1.000000	1.000000	1.000000	0.000000
2	HUNSPELL GERMAN LUCENE FILTER	ALL_CANDIDATES	0.810178	0.331332	0.999994	0.665663	0.999940	0.000060

Pair-relation metrics

Rank	Stemmer	Output policy	F0.5	F1	F2	Jaccard	Fowlkes–Mallows	MCC
1	Radixor	ALL_CANDIDATES	0.998613	0.999133	0.999653	0.998267	0.999133	0.999133
2	HUNSPELL GERMAN LUCENE FILTER	ALL_CANDIDATES	0.628511	0.470321	0.375748	0.307464	0.518110	0.518088

Raw pair counts

Rank	Stemmer	Output policy	TP	FP	FN	TN	Over error / possible	Under error / possible
1	Radixor	ALL_CANDIDATES	858417	1490	0	10594962044	1490 / 10594963534	0 / 858417
2	HUNSPELL GERMAN LUCENE FILTER	ALL_CANDIDATES	284421	66639	573996	10594896895	66639 / 10594963534	573996 / 858417

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	56726	80	1410	10454	7.181227%	8	157137
HUNSPELL GERMAN LUCENE FILTER	10454	18296	7735	4538	3.117315%	3	150205

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: DE_DE
• 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