SIH/lindera-dicts

Lindera Morpheme Dictionaries (JA + KO)

Prebuilt Lindera binary dictionaries for Japanese (IPADIC, UniDic) and Korean (ko-dic) morpheme segmentation, packaged for on-demand download by the LDaCA Text Analytics Tools web app.

The web app uses these dictionaries through its polars-text Rust extension — the first time a user runs tokenisation on a Japanese or Korean corpus, the matching tarball is fetched from this dataset into the per-OS cache directory (~/Library/Caches/ldaca/lindera/ on macOS, ~/.cache/ldaca/lindera/ on Linux, %LOCALAPPDATA%\ldaca\lindera\ on Windows) and reused for every subsequent call. Bundling the dicts in the wheel would push the install size past our hard limit (see memory: project_deployment_targets), so we ship the wheel slim and stream the dict here on first use.

Files

Tarball	Gzipped	Extracted dir	Source dict
ipadic-mecab-2.7.0-bin.tar.gz	~15 MB	ipadic-mecab-2.7.0-bin/	mecab-ipadic 2.7.0-20250920
unidic-mecab-2.1.2-bin.tar.gz	~50 MB	unidic-mecab-2.1.2-bin/	unidic-mecab 2.1.2
ko-dic-2.1.1-bin.tar.gz	~34 MB	ko-dic-2.1.1-bin/	mecab-ko-dic 2.1.1-20180720

Each tarball extracts to a directory containing the standard Lindera binary layout: matrix.mtx, dict.da, dict.vals, dict.words, dict.wordsidx, char_def.bin, unk.bin, metadata.json.

Usage

The dataset is consumed automatically by polars-text when a tokenisation request specifies one of the matching model IDs:

import polars_text as pt
import polars as pl

# First call downloads + caches the dict; subsequent calls hit the cache.
tokens = (
    pl.DataFrame({"text": ["今日は良い天気です"]})
    .lazy()
    .select(pt.tokenize(pl.col("text"), model="lindera-ja-ipadic"))
    .collect()
)
# -> ["今日", "は", "良い", "天気", "です"]

Supported model IDs:

• lindera-ja-ipadic — Japanese, IPADIC dictionary (recommended default)
• lindera-ja-unidic — Japanese, UniDic dictionary (alternative, more accurate for modern web/social text)
• lindera-ko-dic — Korean, ko-dic dictionary

To override the dataset source (for testing or self-hosting), set the LDACA_LINDERA_DICT_REPO environment variable to a different HF dataset path.

Reproduce from source

Each tarball is the output of lindera-cli build against the upstream MeCab source dictionary plus the metadata file shipped in the corresponding Lindera crate (lindera-ipadic, lindera-unidic, lindera-ko-dic):

cargo install lindera-cli --version 3.0.7

# Example for IPADIC:
curl -sSfL -o mecab-ipadic-2.7.0-20250920.tar.gz \
    https://Lindera.dev/mecab-ipadic-2.7.0-20250920.tar.gz
tar -xzf mecab-ipadic-2.7.0-20250920.tar.gz

lindera build \
    --src mecab-ipadic-2.7.0-20250920 \
    --dest ipadic-mecab-2.7.0-bin \
    --metadata <path-to-lindera-ipadic-crate>/metadata.json

tar -czf ipadic-mecab-2.7.0-bin.tar.gz ipadic-mecab-2.7.0-bin/

The upstream MeCab dictionary tarballs have the following MD5s (matching the values recorded in the lindera-ipadic / lindera-unidic / lindera-ko-dic build scripts at version 3.0.7):

• mecab-ipadic 2.7.0-20250920: a95c409f12f1023fce8ef91f991ef042
• unidic-mecab 2.1.2: f4502a563e1da44747f61dcd2b269e35
• mecab-ko-dic 2.1.1-20180720: b996764e91c96bc89dc32ea208514a96

Licensing

Each dictionary inherits the licence of its upstream source. The Lindera build process is a non-creative format conversion (CSV → binary serialisation) and does not introduce new copyrightable content.

• IPADIC — IPA Dictionary Licence (BSD-style, attribution required). Original source: NAIST / Institute of Statistical Mathematics.
• UniDic — BSD-3-Clause / GPL-2 / LGPL-2.1 triple-licensed. Maintained by the National Institute for Japanese Language and Linguistics (NINJAL).
• ko-dic — Apache License 2.0. Originally derived from Sejong corpus data, repackaged by the mecab-ko project.

Lindera itself (the build tool, not the data) is MIT-licensed.

Acknowledgements

This dataset is published by the Sydney Informatics Hub to support text analytics workflows on the Language Data Commons of Australia infrastructure.

Source dictionaries are the work of their respective maintainers:

• mecab / IPADIC — Taku Kudo et al. (Kyoto University, NAIST)
• UniDic — National Institute for Japanese Language and Linguistics (NINJAL)
• mecab-ko-dic — Pusnow / Eunjeon project, building on Sejong Corpus data
• Lindera — Minoru Osuka et al. (the Rust reimplementation that this binary format targets)