Skip to content

Korean stemmer benchmarks ★★★★★

Benchmark-only dictionary — not distributed

ko-kr-default is privately retained for reproducible measurement only. Radixor does not distribute, install, publish, register, or support this dictionary because file-applicable redistribution rights are unresolved. It has no Maven coordinates, BOM/catalog entry, standard-package membership, or Python package.

Evidence: Selected data file lacks file-applicable license evidence. Benchmarking these private bytes does not assert a right to redistribute them.

Dictionary size: ★★★★★. The exact count is 192,673 distinct usable word forms after parser-compatible filtering and exact case-preserved deduplication. Stars rank size among all 167 benchmarked dictionaries; they do not measure linguistic quality, availability, or benchmark accuracy.

The corpus, quality, generalization, and speed evidence below belongs to the private 2026-09-11 Radixor/Java 4.4.0 documentation snapshot. It is never placed on a production or release classpath.

Dictionary corpus

Model ID Private version Language Dictionary rows Distinct usable forms Complete tokens Already-root Changed Timing workload Timing tokens
ko-kr-default 1.0.0 KO_KR 2,672 192,673 196,302 5,344 190,958 changed tokens 190,958

Radixor patch-command distribution

Command class Complete-corpus forms
AppendCharacterCommand 247
BackwardCompoundCommand 166,837
DeleteSuffixCommand 6
PreserveCommand 5,346
ReplaceLastCharacterCommand 23,866

Exact-root accuracy

Stemmer All exact Changed exact Root preserved
Private Radixor filesystem model 99.512% 99.499% 99.963%

Pairwise linguistic quality

OI and UI report over-stemming and under-stemming with raw numerators and denominators. A zero denominator is n/a.

Stemmer Dictionary mode OI raw OI UI raw UI
Private Radixor filesystem model ALL_WORDS 0 / 18554200635 0.000000% 29214 / 7145493 0.408845%
Private Radixor filesystem model LOWERCASE_GROUPS_ONLY 0 / 18554200635 0.000000% 29214 / 7145493 0.408845%

Dictionary-family generalization

This frozen five-seed curve measures transfer to withheld lexical families; it does not establish redistribution or support status.

Training rows Median unseen occurrences Unseen all exact Unseen changed exact Unseen root preserved
10% 176,512 81.085% 80.561% 99.792%
20% 156,840 81.037% 80.512% 99.766%
30% 136,808 82.501% 82.016% 99.786%
40% 117,008 81.947% 81.444% 99.813%
50% 97,497 82.828% 82.353% 99.775%
60% 78,230 83.945% 83.502% 99.719%
70% 58,487 82.242% 81.745% 99.625%
80% 38,779 81.688% 81.167% 99.812%
90% 19,450 82.959% 82.480% 100.000%
100% 0 n/a n/a n/a

Runtime timing

JMH average time uses 3 one-second warmup iterations, 5 one-second measurement iterations, 3 independent forks, and 1 thread.

Benchmark Score ms/op Error ms ns/token
Private Radixor filesystem model 21.239 3.762 111.2