sandpapat/phayathaibert-thai-pos-tagger

PhayaThaiBERT Thai POS Tagger

Fine-tuned PhayaThaiBERT for UPOS Part-of-Speech tagging on Thai sentences.
Trained on the UD_Thai-TUD treebank following Universal Dependencies conventions.

---

Model Description

• Base model: PhayaThaiBERT
• Task: Token Classification (POS tagging)
• Dataset: UD_Thai-TUD
• Tags (15 UPOS):
  ADJ, ADP, ADV, AUX, CCONJ, DET, NOUN, NUM, PART, PRON, PROPN, PUNCT, SCONJ, SYM, VERB

This model predicts word-level POS tags for Thai text. For best performance, use pre-segmented Thai words.

---

Usage

1) Load the model

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("sandpapat/phayathaibert-thai-pos-tagger")
model = AutoModelForTokenClassification.from_pretrained("sandpapat/phayathaibert-thai-pos-tagger")

model.eval()

Option 1: Raw Thai text (automatic word segmentation)

from pythainlp.tokenize import word_tokenize

def predict_pos(text: str):
    """POS tag Thai text (raw string)."""

    # 1. Word segmentation
    words = word_tokenize(text)

    # 2. Tokenize with alignment
    encoded = tokenizer(
        words,
        is_split_into_words=True,
        return_tensors="pt"
    )
    word_ids = encoded.word_ids()

    with torch.no_grad():
        outputs = model(**encoded)
        preds = outputs.logits.argmax(dim=-1)[0]

    # 3. Align subwords → words
    results = []
    prev = None
    for idx, w_id in enumerate(word_ids):
        if w_id is None:
            continue
        if w_id != prev:
            label = model.config.id2label[preds[idx].item()]
            results.append((words[w_id], label))
        prev = w_id

    return results

# Example
text = "ฉันกินข้าวที่ร้านอาหาร"
for w, p in predict_pos(text):
    print(f"{w:15s} {p}")

Option 2: Pre-segmented Words (Recommended - Better Accuracy)

def predict_pos_from_words(words):
    """POS tag a list of pre-segmented Thai words."""

    encoded = tokenizer(
        words,
        is_split_into_words=True,
        return_tensors="pt"
    )
    word_ids = encoded.word_ids()

    with torch.no_grad():
        outputs = model(**encoded)
        preds = outputs.logits.argmax(dim=-1)[0]

    results = []
    prev = None
    for idx, w_id in enumerate(word_ids):
        if w_id is None:
            continue
        if w_id != prev:
            label = model.config.id2label[preds[idx].item()]
            results.append((words[w_id], label))
        prev = w_id

    return results

# Example
words = ["ฉัน", "กิน", "ข้าว", "ที่", "ร้านอาหาร"]
for w, p in predict_pos_from_words(words):
    print(f"{w:15s} {p}")

Example Output

Input: "ฉันกินข้าวที่ร้านอาหาร"

ฉัน: PRON
กิน: VERB
ข้าว: NOUN
ที่: ADP
ร้านอาหาร: NOUN

Training Details

Dataset

• Source: UD_Thai-TUD
• Training Set: 2,902 sentences
• Development Set: 362 sentences
• Test Set: 363 sentences

Training Configuration

• Epochs: 3
• Batch Size: 16
• Learning Rate: 3e-5
• Optimizer: AdamW
• Warmup Ratio: 0.1
• Weight Decay: 0.01

Hardware

• Trained on GPU (CUDA)
• Mixed precision training (FP16)

Limitations

• The model was trained on Universal Dependencies data, which may not cover all domains
• Performance may vary on informal text, social media, or specialized domains
• Word segmentation quality affects accuracy (use consistent segmentation)
• Limited to 15 UPOS tags (coarse-grained POS categories)

Ethical Considerations

• The model should not be used as the sole basis for critical decisions
• Performance may vary across different text types and domains
• Users should validate outputs for their specific use cases

Citation

If you use this model, please cite the base model PhayaThaiBERT:

@inproceedings{lowphansirikul2021wangchanberta,
  title={WangchanBERTa: Pretraining transformer-based Thai Language Models},
  author={Lowphansirikul, Lalita and Polpanumas, Charin and Rutherford, Attapol T and Nutanong, Sarana},
  booktitle={2021 16th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP)},
  pages={1--6},
  year={2021},
  organization={IEEE}
}

Acknowledgements

• Base model: PhayaThaiBERT
• Training data: Universal Dependencies Thai Treebank

License

MIT License'