Context-preserving Sentiment Classification Using Bi-TCN and BI-GRU with Multi-head Self-attention

Abstract

In natural language processing, sentiment classification is the recently used topic. Specifically, the objective of the sentiment analysis is to categorise the polarity expressed on the sentence's target. However, there are some researches for classifying the polarity of the target which outperforms well in their way. Yet, there are some limitations, such as apparent and in-apparent issues, gradient problems, etc., to overcome these issues the context-preserving sentiment classification using BI-TCN (Bidirectional Temporal Convolutional network) and BI-GRU (Bidirectional Gated Recurrent Unit) with Multi-head self-attention is proposed to extracts both the local dependent and global dependent information from the sentence, then it will incrementally extract the supervision information of the target to train the model. Formerly, the model is tested and trained using four datasets and the performance is compared with four existing methods, its accuracy is evaluated using the F1-score, precision, recall, specificity, and MCC (Matthews Correlation Coefficient). Consequently, the proposed approach provides the best accuracy level of 98%.