Bengali Social Media Post Sentiment Analysis using Deep Learning and BERT Model

Social media platforms such as Facebook, Twitter, and others are becoming incredibly popular for expressing sentiments and thoughts. People use these platforms to express not only their happy moments, but also their feelings when they are depressed. Using sentiment analysis in natural language processing to analyse these social media posts, one’s emotional state can be determined, such as happy, sad, or angry at a particular time. The majority of research in this topic is conducted in English, therefore sentiment analysis from Bengali is not very accurate. So, our goal is to work on this topic using Bengali datasets obtained from various social media posts to improve sentiment detection accuracy. This work can be used to help building a system in our country’s mental health sector. In this research, we first gathered social media data. Then we used a number of feature selection and extraction techniques like Word2Vec, GloVe etc. and applied a number of deep learning model, such as RNN, LSTM, GRU etc. We have also applied hybrid and transformer-based BERT models like CNN-BiLSTM Bangla-BERT, mBERT etc and finally got the highest accuracy of 88.59% for the CNN-BiLSTM hybrid model using the GloVe feature vector.