DeepSen: A Deep Learning-based Framework for Sentiment Analysis from Multi-Domain Heterogeneous Data

Abstract

People usually express their emotions, views, or sentiment in textual form. The textual sentiment analysis (TSA) identifies or classifies opinions or feelings from texts in a predefined class. The TSA is complicated or infeasible manually due to its voluminous nature and unstructured or messy conditions. Therefore, the automatic sentiment analysis method quickly paves the way to identify the hidden sentiment polarity from the textual content. Although a few studies on sentiment analysis were conducted on a single or specific domain, developing the TSA method concerning multi-domains is unexplored in Bengali. This paper presents a deep learning-based framework called DeepSen to detect textual sentiment from Bengali texts into three polarities: positive, negative and neutral. Four benchmark corpora from available domains, Book, Restaurant, Drama and Cricket, have been used to analyze sentiment from multi-domain heterogeneous data. This work investigates six popular machine learning (LR, DT, MNB, SVM, RF, AdaBoost) and five deep learning (CNN, LSTM, GRU, BiGRU, BiLSTM) techniques using four benchmark Bengali corpora to perform TSA tasks. The evaluation result reveals that the BiLSTM method obtained the highest weighted f1-score (0.85) among all models.