A Zero-Shot Interpretable Framework for Sentiment Polarity Extraction

• Published in: IEEE access Vol. 12; p. 1
• Main Authors: Chaisen, T., Charoenkwan, P., Kim, C. G., Thiengburanathum, P.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Analytical models; Data mining; Decision analysis; Deep learning; Emotions; Feature extraction; Machine learning; Model Interpretation; Natural language processing; Performance prediction; Predictive models; Semantics; Sentiment analysis; Task analysis; Training; Zero-shot learning
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3322103

Sentiment analysis is a task in natural language processing that focuses on identifying and categorizing emotions expressed in text. Despite the remarkable predictive performance achieved by deep learning models in this domain, their limited interpretability posed a significant challenge. Furthermore, the development of interpretable sentiment analysis models for the Thai language had received inadequate attention. To address this gap, this study proposed a Zero-shot Interpretable Sentiment Analysis Framework, integrating sentiment polarity extraction and leveraging the zero-shot learning with the powerful WangchanBERTa model. Our framework utilized the word selection method from the feeling wheel to represent six primary feelings as sentiment polarities, effectively capturing the subtle emotions expressed in the text. These sentiment polarities played a crucial role as features in training our model, enhancing its interpretability for sentiment analysis tasks. Through the evaluation of three Thai sentiment analysis datasets, we compared the sentiment polarity extraction with two traditional feature extraction methods and ten classification algorithms. The results showed the superiority of the sentiment polarity extraction over Bag of words and its competitive performance compared to TF-IDF in terms of accuracy. To gain insights into the model's decision-making process, SHAP (SHapley Additive exPlanations) was employed to analyze feature importance. Our findings highlighted the alignment of influential features with the sentiment polarities of the text, providing a crucial understanding of the model's functionality. Notably, we uncovered a clear relationship between specific feeling features and their corresponding sentiment classes, deepening our comprehension of the model's performance in sentiment analysis. This study not only contributed to the advancement of sentiment analysis in the Thai language but also bridged the gap between predictive performance and model transparency, yielding a novel and interpretable approach for sentiment analysis.