A Deep Learning Approach for Recognizing Textual Emotion from Bengali-English Code-Mixed Data

Emotion detection is a computational approach for finding the distinct emotion or feeling of an individual. Although Bengali is a low-resource language, the amount of Bengali-English codemixed textual data has grown significantly because of the recent widespread use of social media applications among Bengali users. Gradually, the classification of emotions in Bengali-English codemixed data has become a crucial challenge for applications in e-commerce, healthcare, suicidal attempt reduction and crime detection. Nevertheless, the lack of Bengali language processing techniques and Bengali-English dataset have made the emotion recognition more challenging. This research work offers a Deep Learning based approach for classifying emotions from Bengali-English code-mixed data into six basic categories: disgust, sadness, joy, anger, fear, and surprise. Due to the lack of required dataset, a Bengali-English code-mixed corpus consisting of 10,221 sentences is created. In order to identify the best features, this work investigates several word embedding techniques, including Word2Vec, FastText, and Keras Embedding Layer. Different types of of Machine Learning and Deep Learning based algorithms including the proposed technique using Word2Vec and BiLSTM are applied on the developed corpus. In order to find out the best technique, a comparative analysis among all the methods is demonstrated revealing that the BiLSTM with Word2Vec word embedding technique outperforms rest other models achieving the highest accuracy of 76.1%.