JarBenjaporn/phayathaibert-skincare-absa

PhayaThaiBERT fine-tuned for Thai Skincare/Cosmetic Aspect-Based Sentiment Analysis

This model fine-tunes clicknext/phayathaibert for multi-aspect sentiment analysis of Thai skincare and cosmetic product reviews. It predicts sentiment (positive / negative / neutral / none) independently for 10 aspects in a single forward pass: result/efficacy, price/value, packaging, texture, color, scent, shipping, product packing, seller service, and overall satisfaction.

Trained and evaluated as part of Performance Comparison of Fine-Tuned Thai BERT-based Models for Multi-Aspect Sentiment Analysis in Skincare and Cosmetic Reviews — a comparison against WangchanBERTa, ThaiBERT, and a Typhoon LLM baseline on the same task and dataset. PhayaThaiBERT achieved the best overall performance of the three fine-tuned models.

Dataset: JarBenjaporn/skincare-absa-thai (15,905 Thai reviews from Shopee, Watsons, Konvy)

Architecture: clicknext/phayathaibert encoder + 10 independent linear classification heads (one per aspect), trained jointly with averaged cross-entropy loss.

Hyperparameters & Experimental Setup

Hyperparameter / Config	Tested Values / Details
Base Model Architecture	RoBERTa-base
Base Model Name	pthibert-base-th-cased
Max Sequence Length	256
Optimizer	AdamW
Learning Rate (LR)	1e-5, 2e-5
Batch Size	16, 32
Epochs	3, 5, 7, 9, 11
Dropout	0.1

Best Config:

Learning Rate	Epochs	Batch Size
2e-5	11	16

Test set results:

Metric	Value
Accuracy	0.9412
Macro F1	0.7941
Macro Precision	0.7743
Macro Recall	0.8171
Cohen's Kappa	0.8772
Hamming Loss	0.0588

Comparison across models (test set):

Model	Accuracy	Macro F1	Cohen's κ	Hamming Loss
PhayaThaiBERT	0.9412	0.7941	0.8772	0.0588
WangchanBERTa	0.9354	0.7801	0.8659	0.0646
ThaiBERT (this model)	0.6907	0.382	0.3571	0.3093
Typhoon v2.5 (LLM, no fine-tuning)	0.9393*	0.6679*	0.7388	0.0607

*Typhoon numbers are macro-averaged across aspects from a separate few-shot prompting evaluation (8,123 matched reviews), not directly from the same train/test split — included for context, not a strict apples-to-apples comparison.

Architecture code / how to load: the custom MultiLabelBertForAspectSentiment class is included in this repo (modeling_multilabel_bert.py). Trained weight files are not hosted here due to storage constraints — see the GitHub repo to reproduce training from the dataset.

Limitations: trained on e-commerce reviews from three Thai platforms; performance on other domains (e.g. formal product descriptions, other languages) is not evaluated. Aspect coverage is fixed to the 10 categories above.

---

Credits

Jarbenjaporn, Kronravee