🎉 EMNLP 2025 Workshop Accepted

ThaiFACTUAL

Debiasing Large Language Models in Thai Political Stance Detection via Counterfactual Calibration

A lightweight, model-agnostic framework that mitigates sentiment leakage and political entity bias in Thai political stance detection.

270

Annotated Samples

3

Political Figures

73.5

Macro-F1

55%

Bias Reduction

Overview

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.

Method

Tweet
LLM Prediction
Counterfactual Swap
Rationale Calibration
Debiased Stance

Bias Demonstration

❌ Raw LLM

Positive sentiment → Support

Negative sentiment → Against

Political entity strongly influences prediction.

✅ ThaiFACTUAL

Sentiment separated from stance.

Counterfactual reasoning applied.

More consistent and fair predictions.

Results

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

Paper & Citation

Debiasing Large Language Models in Thai Political Stance Detection via Counterfactual Calibration

@inproceedings{panboonyuen2025thaifactual,
 title={ThaiFACTUAL},
 author={Teerapong Panboonyuen},
 booktitle={EMNLP 2025 Workshop}
}

Author

Teerapong Panboonyuen

Chulalongkorn University

MARSAIL Laboratory