Interpretable model for Thai sentiment analysis using Zero-Shot learning for feature extraction

Abstract

Sentiment analysis is a task that focuses on identifying and categorizing emotions expressed in text. Despite the remarkable predictive performance achieved by deep learning models in this domain, a significant challenge is presented by their limited interpretability. Furthermore, developing interpretable sentiment analysis models for the Thai language has received insufficient attention. To address this gap, a Zero-shot Interpretable Sentiment Analysis Framework is proposed in this study, integrating sentiment polarity extraction and leveraging the zero-shot learning with the WangchanBERTa model. The word selection method from the feeling wheel is employed by our framework to represent six primary feelings as sentiment polarities, capturing the subtle emotions expressed in the text. The sentiment polarities play a crucial role as features in the training of our interpretable model. In evaluating three Thai sentiment analysis datasets, the proposed method demonstrated superiority over Bag of Words, outperforming it by an average of 14.1%. Furthermore, it showcased competitive accuracy compared to TF-IDF, with an average difference of 1.49% across ten classification algorithms. The integration of SHAP (SHapley Additive exPlanations) in our analysis has provided insights into feature importance in the model's decision-making process. Our findings highlight influential features aligning with the sentiment polarities of the text, showing a clear relationship between features and their corresponding sentiment classes. This study improved sentiment analysis in Thai, bridging the gap between predictive performance and model transparency with a novel interpretable approach.