Towards Interpretable Emotion Classification in Bangla: A Hybrid Deep Learning Approach with Explainable AI
Shakib Sadat Shanto, Md. Kishor Morol, Zishan Ahmed, Ahmed Shakib Reza, Md. Abdullah-Al-Jubair · 2024
Emotion classification has emerged as a crucial task in various domains, with significant implications for understanding human behavior and improving human-computer interaction.While Bengali is spoken by a significant population globally, the advancement in developing precise emotion categorization models for Bangla has been relatively limited and insufficient when compared to the strides made in emotion classification for other languages.To address this gap, we propose a novel hybrid deep neural network (DNN) architecture, BI-GRU+LSTM+CNN, specifically designed for emotion classification in Bangla social media comments.Our approach integrates the Local Interpretable Model-Agnostic Explanations (LIME) technique to provide interpretability to the model's predictions, enabling the identification of the most influential words and phrases driving the classification of each emotion.In order to assess the efficiency of our proposed model, we performed tests on a substantial dataset consisting of 42,604 Bangla social media comments.This dataset encompasses six unique emotion categories.The proposed BI-GRU+LSTM+CNN architecture achieved a remarkable accuracy of 90.71%, surpassing traditional machine learning models and other deep learning architectures in the task of Bangla emotion classification.Moreover, our approach outperforms existing six-class emotion classifiers for Bangla, demonstrating its superior ability to capture the intricate nuances of emotions in Bangla.