HAN-CapsNet: Multi-polarity Based Sentiment Classification using Hierarchical Attention Networks and Capsule Neural Network

Aspect-based Sentiment classification intends to find sentiment polarity towards a particular aspect becoming a significant task for sentiment analysis. Conventional techniques obtained satisfactory outcomes, yet they mainly focused on utilizing local structure information of the given sentence, like syntactical dependency limitations in a sentence. To solve this issue, the Hierarchical Attention Networks-Capsule Neural Network (HAN-CapsNet) is proposed for multi-polarity based sentiment classification. Initially, BERT tokenization is performed using an input text from a database. The second task considers BiLSTM for Multi-Word Aspect Term (MWAT) extraction. Later, feature extraction is done where the extracted features are statistical, count vectorization, emoticons, hashtags, all caps, elongated units, negation, bag-of-units, syntactic n-gram features, and contextual features. Finally, MWAT-based polarity sentiment classification is accomplished by utilizing HAN-CapsNet. The BiLSTM outperforms previous feature extraction algorithms with an accuracy rise of 3% - 6%. Moreover, the performance measures of the proposed model for sentiment classification, like Accuracy, True Positive Rate, and True Negative Rate show significant improvements over the baseline models.