Croatian Stemmer

A rule-based stemmer for Croatian, available as a Python library and as a PostgreSQL text search dictionary. Both implementations produce identical stems.

It starts from the Ljubešić–Pandžić stemmer ("Simple stemmer for Croatian v0.1", 2012), which refines the algorithm from Ljubešić, Boras & Kubelka (2007). On top of that it adds extra suffix rules and post-processing for comparatives, jotation and fleeting a. These additions were tuned against the hr500k corpus and checked on a separate held-out corpus; see Evaluation.

Python

Requires Python 3.10 or newer.

pip install croatian-stemmer

from croatian_stemmer import stem, stem_text

stem("gradovima")   # "grad"
stem("kuće")        # "kuć"
stem("lijepog")     # "lijep"
stem("metara")      # "metr"   (fleeting a)
stem("najvećih")    # "već"    (superlative prefix)

stem_text("Veliki grad ima lijepe kuće")
# ["velik", "grad", "im", "lijep", "kuć"]

stem_text("Ja sam bio u gradu")
# ["ja", "u", "grad"]   (stop words "sam" and "bio" are dropped)

stem() stems any single word. stem_text() tokenizes, lowercases and stems running text, dropping the 67 stop words: forms of the auxiliaries and modals biti, htjeti, željeti, morati, trebati and moći. Passing a stop word to stem() returns it unchanged (stem("sam") == "sam").

Command line

croatian-stemmer kuća grada lijepog        # word<TAB>stem per line
echo "gradovima" | croatian-stemmer        # one word per line on stdin
echo "Ja sam bio u gradu" | croatian-stemmer -t   # free text: ja u grad

PostgreSQL extension

pg/ contains a PGXS extension, dict_croatian_stem, tested on PostgreSQL 15, 16, 17 and 18. The stemmer is plain C with no dependencies. Its rule tables are generated from the Python rules by gen_rules_c.py, and the test suite checks that the C and Python stems match.

Build and install

Requires the server development headers (postgresql-server-dev-NN on Debian/Ubuntu).

cd pg/
make USE_PGXS=1
sudo make USE_PGXS=1 install

To build and run the regression tests in Docker against several versions:

cd pg/
./docker-test.sh          # PG 15, 16, 17, 18
./docker-test.sh 17 18    # chosen versions

Usage

The extension is marked trusted = true, like contrib's unaccent, so once it is installed on the server any user with CREATE privilege on a database, such as its owner, can run CREATE EXTENSION dict_croatian_stem without superuser. If you do not want that, delete the trusted line from dict_croatian_stem.control before installing.

CREATE EXTENSION dict_croatian_stem;
-- Creates the croatian_stem dictionary, the croatian_stem_template
-- template, and the croatian text search configuration.

SELECT to_tsvector('croatian', 'Veliki grad ima lijepe kuće');
-- 'grad':2 'im':3 'kuć':5 'lijep':4 'velik':1

SELECT to_tsvector('croatian', 'Lijepe kuće u velikom gradu')
    @@ to_tsquery('croatian', 'kuća & grad');
-- t

SELECT ts_lexize('croatian_stem', 'gradovima');   -- {grad}
SELECT ts_lexize('croatian_stem', 'sam');         -- {}  (stop word)

The dictionary always drops the built-in stop words. To drop more, create a dictionary from the template with your own list in $SHAREDIR/tsearch_data/<name>.stop:

CREATE TEXT SEARCH DICTIONARY my_croatian_stem (
    TEMPLATE = croatian_stem_template,
    StopWords = my_croatian
);

How it works

1. Stop words. The 67 auxiliary and modal forms are returned unchanged.
2. Superlatives. A leading naj- is peeled off, so najveći, veći and velik can meet.
3. Transformations. 159 suffix rewrites undo consonant alternations (palatalization, sibilarization, jotation) and fleeting a.
4. Suffix stripping. 81 ordered rules, each a stem pattern plus a set of suffixes. The first rule that leaves a stem with a vowel and more than one character wins. Syllabic r (trg, krv, prst) counts as a vowel.
5. Verbs. The thematic vowel is dropped the same way in every form, so raditi, radim, radila and radi all give rad, and vidjeti, vidio and vidi give vid. -ovati/-ivati verbs and their -uj- present share a stem (djelovati, djeluje → djeliv), and the -ići compounds keep their đ (došao, dođe → dođ).
6. Participles. Passive participles lose -en, -an, -iran or -nut and undo the root alternation, so they meet their verb (objavljen → objav, korišten → korist, organiziran → organizir).
7. Clean-up. Fleeting a (posao/posla, postupak/postupka, kupac/kupca, metar/metra), the -ov-/-ev- plural extension and comparative -ij are normalized so related forms share a stem.

Evaluation

The rules were tuned on hr500k (CLARIN.SI, CC BY-SA 4.0; 427,270 word tokens with gold lemmas and parts of speech), so hr500k numbers are in-sample. The held-out check is ParlaMint-HR 4.1 (CLARIN.SI, CC BY 4.0): Croatian parliamentary debates from 2003 to 2022, 85.4 million word tokens, lemmatized and tagged automatically by CLASSLA-Stanza. None of it was looked at during tuning.

The 67 stop words are left out of the scoring, since stem() returns them unchanged and text search drops them.

The v0.1 hr500k figures (78.4% form-level consistency, 8.6% over-stemming) differ slightly from docs/evaluation.md (78.3%, 8.7%) because the stop words are now left out of the scoring.

The gains over v0.1 hold on the held-out corpus, so the rules generalize rather than fit hr500k. Much of the added over-stemming is a verb meeting its own noun or participle (dokaz/dokazati, određen/odrediti), which the corpora lemmatize separately but which is usually what search wants. Over-stemming grows with vocabulary size (ParlaMint has six times as many distinct form/lemma pairs as hr500k), so compare it between stemmers on the same corpus, not across corpora. An evaluation against human judgments is in progress; see docs/labeling.md.

An earlier version also carried a hand-built lexicon of about 3,800 verb and adjective forms chosen from hr500k failures. It raised hr500k verb consistency from 28.3% to 36.3% but ParlaMint-HR verb consistency only from 22.8% to 23.1% (both measured with the stop words included), so it was removed.

To reproduce, download hr500k.conllup into data/ and run:

python evaluate.py          # full hr500k report
python evaluate_score.py    # the composite score used during tuning

# held-out: stream ParlaMint-HR.ana.tgz from the link above
tar -xzOf ParlaMint-HR.ana.tgz '*.conllu' | python evaluate_score.py --corpus -

docs/evaluation.md is the detailed report for the v0.1 baseline.

How the rules were tuned

The rules were tuned by an automated agent working in a loop. AUTORESEARCH.md is the brief it was given, and results.tsv is its experiment log, one row per experiment. The loop only ever saw hr500k; ParlaMint-HR was the held-out check, run between loops. Per-word lists (lexicons, exception tables, extra stop words) are banned, so every change is a general rule or algorithm change.

Development

pip install -e ".[dev]"
pytest                      # Python tests, plus C/Python parity if a C compiler is present
python gen_rules_c.py       # regenerate pg/croatian_stemmer_rules.h after editing rules.py

CI regenerates the C rule header and fails if it differs from the committed one.

Repository layout

src/croatian_stemmer/      Python package (stemmer.py, rules.py, CLI)
src/croatian_stemmer/labeling/  optional human-labeling tools and web app
tests/                     pytest suite, including C/Python parity
pg/                        PostgreSQL extension, C stemmer and regression tests
gen_rules_c.py             generates pg/croatian_stemmer_rules.h from rules.py
evaluate.py, evaluate_score.py   corpus evaluation
AUTORESEARCH.md            brief for the automated rule-tuning agent
results.tsv                tuning experiment log, one row per experiment
docs/                      research notes, evaluation reports, labeling guides
build.sh                   builds and pushes the labeling app image
Dockerfile.web             container image for the labeling app
migrations/, alembic.ini   labeling app database migrations (Alembic)
scripts/                   labeling app asset build and test runner
package.json               labeling app JS dependencies (Keycloak adapter, Playwright)
requirements-web.txt       hashed, pinned dependencies for the labeling container

The labeling app (a public, multi-user site: Keycloak/Google sign-in, consent gate, shared curator-published corpora, judgment forms, exports) is optional and is not needed to use the stemmer. It runs from a checkout (pip install -e ".[web]") or the container built from Dockerfile.web, not from the PyPI package, which ships only the offline croatian-stemmer-labels tools. See docs/labeling-web.md.

References

• Ljubešić, N., Boras, D., & Kubelka, O. (2007). Retrieving Information in Croatian: Building a Simple and Efficient Rule-Based Stemmer. In INFuture2007: Digital Information and Heritage, 313–320. Zagreb.
• Ljubešić, N., & Pandžić, I. (2012). Simple stemmer for Croatian v0.1. http://nlp.ffzg.hr/resources/tools/stemmer-for-croatian/
• Ljubešić, N., et al. hr500k, a reference training corpus of Croatian, v2.0. CLARIN.SI. http://hdl.handle.net/11356/1792

License

GNU Lesser General Public License v3.0 or later; see COPYING.LESSER and COPYING. The stop list, transformations and suffix rules come from "Simple stemmer for Croatian v0.1", Copyright 2012 Nikola Ljubešić and Ivan Pandžić, also LGPL-3.0-or-later.

The hr500k corpus is not included. It is licensed CC BY-SA 4.0 by its authors.