Sentiment Polarity Detection of Book Reviews in Bangla Using Deep Neural Network

Abstract

Sentiment analysis is a Natural Language Processing (NLP) task that extracts people’s feelings or opinions from words or documents. Sentiment analysis has already found acceptance in rich languages such as English. Unfortunately, research in this topic is very limited for Bangla. The majority of the researchers employed standard machine learning techniques with unimpressive results. Some researchers used deep learning techniques, but their accuracy was sub-par. Most of them ran into problems due to a lack of sufficient Bangla datasets. So, our main objective was improving sentiment polarity detection model as well as it’s accuracy. We proposed a deep learning based method for detecting sentiment polarity in the Bangla Book Review in this research. Our model is a hybrid architecture composed of recurrent and convolutional layers. The model’s dataset was correctly pre-processed by using tag removal, word token generation, POS tagging, punctuation removal and tokenization. It improved the model’s effectiveness and performance. This algorithm achieved a greater accuracy of 92.54% whereas the precision is 93.07%, recall is 92.16% and F1 score is 92.61%.