An Explainable AI Framework for Bias Detection in Indonesian Sentiment Analysis

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The development of deep learning models in sentiment analysis has significantly improved the accuracy of text classification. However, these models are generally black-box, making them difficult to interpret and potentially contain undetectable biases. This study aims to detect and evaluate bias in sentiment analysis models using SHAP-based Explainable Artificial Intelligence (XAI) approach. The model used is IndoBERT which is trained on the Indonesian sentiment dataset. The research stages include data preprocessing, model training, interpretability analysis using SHAP, and fairness evaluation using disparate impact metrics. The results showed that the model achieved an accuracy of 88%, but SHAP analysis revealed the dominance of certain features in decision-making. In addition, fairness evaluation shows that there is an inequality in the distribution of predictions between groups. These findings show that high accuracy does not guarantee model fairness, so Explainable AI integration is important in the development of transparent and accountable NLP systems.