Datasets:

Modalities:
Text
Formats:
json
Languages:
Thai
ArXiv:
License:
yacht's picture
Add the arXiv number of the paper (2611.21300) to the links and the citation.
08fb0d9 verified
|
Raw History Blame Contribute Delete
9.86 kB
metadata
language:
  - th
license:
  - cc-by-sa-4.0
  - cc-by-sa-3.0
  - cc0-1.0
  - apache-2.0
pretty_name: OpenThai2.0 WordSeg Benchmark
task_categories:
  - token-classification
tags:
  - thai
  - word-segmentation
  - tokenization
  - benchmark
size_categories:
  - 10K<n<100K
configs:
  - config_name: tud
    default: true
    data_files:
      - split: test
        path: data/tud.jsonl
  - config_name: pud
    data_files:
      - split: test
        path: data/pud.jsonl
  - config_name: ws160
    data_files:
      - split: test
        path: data/ws160.jsonl
  - config_name: vistec
    data_files:
      - split: test
        path: data/vistec.jsonl

OpenThai2.0-WordSeg-Benchmark

A standard test for Thai word segmentation, version 1.0: seven test sets, one scoring rule, and the results of twelve systems. It accompanies the LATTE-WordSeg-Lite paper.

Quick start

pip install -U huggingface_hub
hf download iapp/OpenThai2.0-WordSeg-Benchmark --repo-type dataset --local-dir wordseg-bench
cd wordseg-bench && pip install -r requirements.txt
from benchmark import evaluate

evaluate(my_segmenter, name="my-segmenter")   # my_segmenter(text) -> list of words

This runs your segmenter on the 40,073 lines of the seven test sets. It prints word F1 and boundary F1 on each set and their averages, next to the published systems, and saves results-my-segmenter.json. From the command line: python benchmark.py --module my_package:segment, or python benchmark.py --predictions my_outputs/ with one file per set, SET.jsonl, holding {"id": ..., "tokens": [...]} for each line of data/SET.jsonl.

Why the first run downloads files. Four test sets are in this repository. The other three (BEST, ORCHID and khammoon's test) have licences that do not allow sharing them. The first run downloads them from their owners (about 50 MB), checks them and builds them here, once.

Test sets

Set Source Licence Included
tud UD Thai-TUD, test split CC BY-SA 4.0 yes
pud UD Thai-PUD CC BY-SA 3.0 yes
ws160 Wisesight-160 CC0 1.0 yes
vistec VISTEC-TP-TH-2021, test split CC BY-SA 3.0 yes
bestdd BEST 2009, our split CC BY-NC-SA 3.0 no
orchidndd ORCHID, our split CC BY-NC-SA 3.0 no
khm khammoon data, test split CC BY-NC 4.0 no

How the sets were made from their sources is listed in NOTICE and explained in the paper.

Results

Average F1 (%) over the seven test sets:

System Word F1 Boundary F1
LATTE-WordSeg-Lite 85.6 93.3
SEFR CUT, best 84.5 92.6
DeepCut 84.5 92.6
AttaCut-SC 84.0 92.4
SEFR CUT, ws1000 84.0 92.5
AttaCut-C 83.2 91.9
khammoon 81.8 91.3
OSKut, scads 79.4 89.9
TLTK 75.5 87.9
OSKut, ws 75.3 88.1
PyThaiNLP nlpO3 70.9 85.6
PyThaiNLP newmm 70.9 85.7

LATTE-WordSeg-Lite with its default unit, short (long: 83.1 and 92.0, minimal: 82.1 and 91.5). khammoon segments morphemes, not words: six of the seven sets count words, so compare it with minimal. Boundary F1 scores each cut between two characters, so it is higher than word F1 and closer between systems. Both scores on each set are in results/expected.json, and every system's output is in predictions/. The model is at iapp/OpenThai2.0-LATTE-WordSeg-Lite and in the Python package iapp-ai.

How scores are computed

  • Word F1: a word is correct when a gold word has the same start and end. Boundary F1: a cut between two characters is correct when gold cuts there too. A space always separates words and is never scored: a token that holds a space counts as the words between its spaces.
  • Every system gets the same lines and the same rule, so its numbers can be compared with the table above.

Data format

Each line of data/SET.jsonl is one test line, with the start and end of every gold word:

{"id": 0, "text": "ยุ้ยจะบอกว่ามันนานแล้ว แต่เรารักกัน", "gold": [[0, 4], [4, 6], [6, 9], [9, 12], [12, 15], [15, 18], [18, 22], [23, 26], [26, 29], [29, 32], [32, 35]]}

The four included sets also load with datasets: load_dataset("iapp/OpenThai2.0-WordSeg-Benchmark", "tud").

Reproduce the paper

python fetch_sources.py     # every original file at its pinned revision, SHA-256 checked
python build_testsets.py    # every test file, checked against checksums.json
python score.py table       # every score of the paper, checked against results/expected.json

Licences

  • Test files: the licence of their source (table above). Files built from a CC BY-SA corpus are shared under the same licence. NOTICE gives the attribution of each source.
  • Code (*.py): Apache-2.0 (LICENSE).
  • Predictions: the word offsets contain no text; the token files of the four included sets contain their text, under its licence.

Citation

If you use the benchmark, please cite the paper and the source of each test set you use.

@misc{chayintr2026lattewordseglite,
  title         = {{LATTE-WordSeg-Lite}: Multi-Granularity {Thai} Word Segmentation on a {CPU}, with a Reproducible Benchmark},
  author        = {Chay-intr, Thodsaporn and Viriyayudhakorn, Kobkrit and Theeramunkong, Thanaruk},
  year          = {2026},
  eprint        = {2611.21300},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2611.21300}
}
BibTeX of the sources
@article{sriwirote2025tud,
  title   = {The {Thai} {Universal} {Dependency} Treebank},
  author  = {Sriwirote, Panyut and Leong, Wei Qi and Polpanumas, Charin and Thanyawong, Santhawat and Tjhi, William Chandra and Aroonmanakun, Wirote and Rutherford, Attapol T.},
  journal = {Transactions of the Association for Computational Linguistics},
  volume  = {13},
  pages   = {376--391},
  year    = {2025},
  doi     = {10.1162/tacl_a_00745}
}

@inproceedings{zeman2017conll,
  title     = {{CoNLL} 2017 Shared Task: Multilingual Parsing from Raw Text to {Universal Dependencies}},
  author    = {Zeman, Daniel and others},
  booktitle = {Proceedings of the {CoNLL} 2017 Shared Task: Multilingual Parsing from Raw Text to {Universal Dependencies}},
  pages     = {1--19},
  year      = {2017}
}

@misc{suriyawongkul2024wisesight,
  title     = {{PyThaiNLP/Wisesight} Sentiment Corpus with Word Tokenization Label},
  author    = {Suriyawongkul, Arthit and Chuangsuwanich, Ekapol and Chormai, Pattarawat and Chantarapratin, Nitchakarn and Prasertsom, Ponrawee and Sawatphol, Jitkapat and Yamada, Nozomi and Rutherford, Attapol and Polpanumas, Charin and Udomcharoenchaikit, Can},
  publisher = {Zenodo},
  note      = {Version 1.1},
  doi       = {10.5281/zenodo.3457446},
  year      = {2024}
}

@inproceedings{limkonchotiwat2021oskut,
  title     = {Handling Cross- and Out-of-Domain Samples in {Thai} Word Segmentation},
  author    = {Limkonchotiwat, Peerat and Phatthiyaphaibun, Wannaphong and Sarwar, Raheem and Chuangsuwanich, Ekapol and Nutanong, Sarana},
  booktitle = {Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021},
  pages     = {1003--1016},
  year      = {2021}
}

@inproceedings{kosawat2009best,
  title     = {{BEST} 2009: {Thai} Word Segmentation Software Contest},
  author    = {Kosawat, Krit and Boriboon, Monthika and Chootrakool, Patcharika and Chotimongkol, Ananlada and Klaithin, Supon and Kongyoung, Sarawoot and Kriengket, Kanyanut and Phaholphinyo, Sitthaa and Purodakananda, Sumonmas and Thanakulwarapas, Tipraporn and Wutiwiwatchai, Chai},
  booktitle = {Eighth International Symposium on Natural Language Processing},
  pages     = {83--88},
  year      = {2009}
}

@article{sornlertlamvanich1999orchid,
  title   = {Building a {Thai} Part-of-Speech Tagged Corpus ({ORCHID})},
  author  = {Sornlertlamvanich, Virach and Takahashi, Naoto and Isahara, Hitoshi},
  journal = {Journal of the Acoustical Society of Japan (E)},
  volume  = {20},
  number  = {3},
  pages   = {189--198},
  year    = {1999},
  doi     = {10.1250/ast.20.189}
}

@article{khammoon2026,
  title   = {khammoon: From a Linguistic Guideline to a {Thai} Morpheme Segmenter},
  author  = {Sirichotedumrong, Warit and Na-Thalang, Adisai and Wittayasakpan, Chanakan and Phatcharoen, Kritsadha and Pipatanakul, Kunat},
  journal = {arXiv preprint arXiv:2611.13012},
  year    = {2026}
}

@inproceedings{limkonchotiwat2025wangchan,
  title     = {{WangchanThaiInstruct}: An Instruction-Following Dataset for Culture-Aware, Multitask, and Multi-domain Evaluation in {Thai}},
  author    = {Limkonchotiwat, Peerat and Tuchinda, Pume and Lowphansirikul, Lalita and Nonesung, Surapon and Tasawong, Panuthep and Aji, Alham Fikri and Udomcharoenchaikit, Can and Nutanong, Sarana},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  pages     = {3535--3558},
  year      = {2025}
}