Dictionary-trained. Not dictionary-bound.

High-quality multilingual stemming.
Built for Java and Python.

Radixor learns word-to-stem transformations from lexical evidence, compiles them into compact tries of patch commands, and applies those transformations algorithmically at runtime. It combines deterministic results, small deployable models, and high throughput across 143 language defaults. The training dictionary supplies evidence—it does not define a closed runtime vocabulary.

Choose your runtime

♨Get Radixor for Java → PyChoose a Python runtime →

Learned transformationsTrie + compact patch commands

GeneralizesNot restricted to listed word forms

ReproduciblePublic quality and speed evidence

Why Radixor is different → Language benchmarks → Python benchmarks → Scientific results & publications → Choose a documentation path →

Measured Python runtime performance

31 / 31Python standard models
benchmarked

29 / 29Comparisons where both Python
runtimes beat PyStemmer

2.165×PyO3 geometric-mean
speedup vs PyStemmer

2.278×Python-C geometric-mean
speedup vs PyStemmer

Python (PyO3) performance (batch size N=100)

6.29M – 15.68M words per second

20M15M10M5M0

6.29MSlowest15.68MFastest

* The 2026-09-11 Python batch report contains 29 direct PyStemmer 3.1.0 (Snowball C) comparisons. Radixor 4.4.0, with the 31-model `radixor-models-standard` 3.0.0 package, records lower median processing time in 29/29 comparisons through PyO3 and 29/29 through Python-C. The geometric-mean speedups are 2.165× and 2.278× respectively. Each ratio is PyStemmer ns/word divided by Radixor ns/word, so values above 1 favor Radixor. PyStemmer has no direct Hebrew or Ukrainian comparator. The chart shows PyO3 throughput; Python-C spans 6.44M–17.81M words/s. View the complete Python benchmarks →

Java reference agreement and Python runtime — Finnish case study

Radixor❄ Snowball Finnish Radixor advantage

Balanced accuracyJava quality evaluation · higher is better0.980.74+33.11% higher

◌Under-stemmingJava quality evaluation · lower is better3.03%52.03%94.17% lower

Python PyO3 speedbatch N=100 · higher ratio is faster1.26× vs PyStemmer1.00× PyStemmer baseline1.26× faster

Java same-resource agreement: balanced accuracy and under-stemming use a Finnish model trained from the lexical resource that defines the reference relation. This is not independent external linguistic validation, and no Python quality evaluation is implied. Python runtime: PyO3 measured 6.73M words/s versus PyStemmer 3.1.0 at 5.36M words/s (1.26×), and Python-C measured 1.49×, all at batch size N=100.

144 models · 143 language pages

♨Java

import org.egothor.stemmer.StemmerPatchTrieLoader;
import org.egothor.stemmer.FrequencyTrie;

FrequencyTrie<CompiledPatchCommand> trie =
    StemmerPatchTrieLoader.loadCompiled(
        StemmerPatchTrieLoader.Language.US_UK, ...);

String word = "running";
CompiledPatchCommand patch = trie.get(word);
String stem = patch == null ? word : patch.apply(word);

View Java docs →

PyPython

from radixor import Stemmer

stemmer = Stemmer("en")
stems = stemmer.stem_batch([
    "running", "studies", "better", "cars"
    ])

View Python docs →

Learned transformationsLexical evidence becomes
compact patch commands

Beyond the dictionaryRuntime is not a closed
word-to-lemma lookup

Native speedMulti-million words per
second in Python and Java

144 model artifactsOne architecture, trained
from language-specific data

Production readyDeterministic behavior
and reproducible evidence

A different stemming architecture.

Radixor combines learned patch commands, trie-based structural sharing, subtree reduction, and deterministic runtime application. It is neither a flat dictionary lookup nor another fixed suffix-rule table.

Technology & lineage ▣ Read the Docs

23 benchmark-only dictionariesNot distributed or installable; license evidence unresolved

High-quality multilingual stemming.Built for Java and Python.

Java reference agreement and Python runtime — Finnish case study

A different stemming architecture.

High-quality multilingual stemming.
Built for Java and Python.