phayathaibert-thai-pos-tagger-onnx

This is an ONNX version of nlp-chula/phayathaibert-thai-pos-tagger, a PhayaThaiBERT model fine-tuned for Universal POS tagging on UD Thai-TUD. It runs with onnxruntime and tokenizers, without PyTorch or Transformers.

This repository only converts the model. All credit for training goes to the original authors at nlp-chula and to the PhayaThaiBERT authors.

Made for the PyThaiNLP POS tagging engine (issue #1536).

Files

Usage

Input is a list of pre-segmented Thai words. Each word is tagged with the prediction for its first subword, the same alignment the model was trained with.

pip install onnxruntime tokenizers numpy huggingface_hub

import json

import numpy as np
import onnxruntime as ort
from huggingface_hub import snapshot_download
from tokenizers import Tokenizer

model_dir = snapshot_download("wiriyabot/phayathaibert-thai-pos-tagger-onnx")
session = ort.InferenceSession(f"{model_dir}/model.onnx", providers=["CPUExecutionProvider"])
tokenizer = Tokenizer.from_file(f"{model_dir}/tokenizer.json")
id2label = {int(k): v for k, v in json.load(open(f"{model_dir}/config.json"))["id2label"].items()}

words = ["ฉัน", "กิน", "ข้าว", "ที่", "ร้านอาหาร"]
enc = tokenizer.encode(words, is_pretokenized=True)
logits = session.run(None, {
    "input_ids": np.array([enc.ids], dtype=np.int64),
    "attention_mask": np.array([enc.attention_mask], dtype=np.int64),
})[0][0]

tags, prev = [], None
for i, word_id in enumerate(enc.word_ids):
    if word_id is not None and word_id != prev:
        tags.append(id2label[int(logits[i].argmax())])
    prev = word_id
print(list(zip(words, tags)))
# [('ฉัน', 'PRON'), ('กิน', 'VERB'), ('ข้าว', 'NOUN'), ('ที่', 'ADP'), ('ร้านอาหาร', 'NOUN')]

Input length

Keep each model call at 415 subwords or fewer, counting <s> and </s>. Split longer input into sentences, or into chunks of whole words, and tag each one separately.

• The model runs on up to 510 tokens (longer input fails), but only the first 415 are reliable. From position 415 on, accuracy drops from about 89% to about 34%. The original PyTorch model does the same, so the drop is not caused by the ONNX export. It suggests that the model was pretrained on sequences of about 416 tokens.
• tokenizer.json truncates at 510 tokens without a warning. This setting comes from the original model. Words after the cut get no subword, so the loop above returns fewer tags than words. Call tokenizer.no_truncation() and check the length yourself.

Measured on the UD Thai-TUD test set, tagged as one long text in chunks of whole words:

The PyThaiNLP engine pos_tag(words, engine="phayathaibert") (issue #1536) chunks long input this way.

Evaluation

All results below are on the UD Thai-TUD test set: 363 sentences, 7,683 words, gold segmentation. Each run was on CPU (Intel Core i5-1340P) with onnxruntime 1.30, using scripts/compare_with_pytorch.py.

• Speed. ONNX averaged roughly 35–45 ms per sentence on this CPU; PyTorch averaged about 120 ms.
• Batching. Outputs are identical whether a sentence is run alone or inside a padded batch.
• Tokenizer. The tokenizers encoding matches the Transformers tokenizer on every test sentence.
• Which weights are int8. Only the embedding tables (Gather: word, position and token type embeddings) are quantized. They hold 69% of the weights and tolerate int8 well. Quantizing the attention and feed-forward MatMul layers too costs 3.3 accuracy points. It also makes the results depend on the batch, so those layers stay float32.
• Comparison with the model card. The original model card reports 90.64%. The higher figure here was measured with the same alignment, by a different evaluation script.

Reproduce

pip install torch transformers onnx onnxruntime
hf download nlp-chula/phayathaibert-thai-pos-tagger --local-dir source
python scripts/export_onnx.py source out          # export + embedding quantization
python scripts/pos_tag_onnx.py out th_tud-ud-test.conllu

Citation

Please cite PhayaThaiBERT:

@article{Sriwirote_2025,
  title={PhayaThaiBERT: Enhancing a Pretrained Thai Language Model with Unassimilated Loanwords},
  volume={24},
  ISSN={2375-4702},
  url={http://dx.doi.org/10.1145/3765962},
  DOI={10.1145/3765962},
  number={11},
  journal={ACM Transactions on Asian and Low-Resource Language Information Processing},
  publisher={Association for Computing Machinery (ACM)},
  author={Sriwirote, Panyut and Rutherford, Attapol T. and Thapiang, Jalinee and Timtong, Vasan},
  year={2025},
  month=Oct,
  pages={1–17}
}

License

MIT, the same as the original model. Training data: the UD Thai-TUD treebank (CC BY-SA 4.0).