Bengali Human Sentiment Detection and Classification using Attention based Bidirectional LSTM

Sentiment Analysis has become of great importance due to the rising volumes of user-generated content on diverse social media and online platforms in Bengali text. However, correctly assessing the sentiment behind Bengali text presents unique challenges because of its complex morphology, variety of dialects, and ungrammatical mixture of code. To overcome this limitation, a bidirectional LSTM with a self-attention-based approach is proposed to detect six sentiments in Bengali text. The proposed model is indeed very forward in capturing long-distance dependencies from both directions in the past and future. The BiLSTM model enhances sentiment detection by being able to learn long-range dependencies in both directions, and the attention mechanism enhances the performance by emphasizing sentiment-bearing words in the text. This model is trained on our own dataset, which comprises 5000 Bengali texts, obtaining almost 85% accuracy. The result is empirical proof of the effectiveness of our method compared to the current state of the art. Presented results would enable contributing for least resources acting as a stage for more sentiment detection work in Bengali.