부산대학교 도서관

Emotion classification holds significant importance in various domains. However, the development of accurate emotion classification models for the Bengali language has been relatively limited, despite its vast speaker base. The unique characteristics of Bengali present several challenges for emotion classification. Consequently, there is an urgent demand for robust and contextually-aware emotion classification models tailored to the linguistic nuances of Bengali. This paper presents a comprehensive study on emotion classification in Bengali text, aiming to develop robust and effective models specific to the language. We explored a range of ML and DL models, including LR, SVC, CNN, LSTM, and BiLSTM. Additionally, we proposed novel hybrid architectures, combining CNN with LSTM and CNN with BiLSTM, to leverage both local and contextual information from Bengali text. However, the lack of comprehensive Bengali emotion datasets further hinders the development of dedicated emotion classification models for the language. To facilitate research, we created the ‘Bengali Emotion Dataset’ consisting of 14,334 social media comments, accurately labeled into seven emotion classes. The results demonstrate that the hybrid models, particularly CNN+BiLSTM, outperform individual ML and DL models, achieving the highest accuracy and F1 score of 88.45% and 88.42% respectively. The benchmark dataset and the success of the hybrid model pave the way for more empathetic and contextually-aware natural language processing applications for Bengali speakers.