A lightweight, model-agnostic framework that mitigates sentiment leakage and political entity bias in Thai political stance detection.
Annotated Samples
Political Figures
Macro-F1
Bias Reduction
Thai political discourse presents unique challenges for LLMs due to sentiment-stance entanglement, indirect language, and entity preference bias. ThaiFACTUAL introduces counterfactual calibration to disentangle political stance from emotional tone.
Positive sentiment → Support
Negative sentiment → Against
Political entity strongly influences prediction.
Sentiment separated from stance.
Counterfactual reasoning applied.
More consistent and fair predictions.
| Model | Bias ↓ | F1 ↑ | OOD ↑ |
|---|---|---|---|
| GPT-4 | 21.7 | 70.8 | 56.4 |
| GPT-4 Debias Prompt | 18.3 | 71.9 | 57.0 |
| LLaMA-3 CoT | 16.5 | 68.1 | 59.7 |
| ThaiFACTUAL | 9.8 | 73.5 | 65.2 |
Debiasing Large Language Models in Thai Political Stance Detection via Counterfactual Calibration
@inproceedings{panboonyuen2025thaifactual,
title={ThaiFACTUAL},
author={Teerapong Panboonyuen},
booktitle={EMNLP 2025 Workshop}
}