Aspect-Based Sentiment Analysis Using BiLSTM-GateCNN Architecture

Abstract

Conducting sentiment analysis on a large amount of comment text is an important means of extracting potential information from the text. However, existing aspect-based sentiment analysis tasks face difficulties in effectively capturing the associations between words and thus cannot accurately extract aspect words, they also do not fully utilize contextual semantic features, which overlooks the connection between aspect words and the context. This study proposes an aspect-based sentiment analysis model based on the BiLSTM-GateCNN network. First, we integrate the self-attention mechanism into the BiLSTM layer to obtain sentence feature representations and aspect term information. Then, we use the GateCNN network with multiple filters in the convolution layer to effectively extract feature information for multiple aspects in each receiving field. Finally, the gated unit simultaneously generates aspect features and sentiment features, making the text more discriminative in emotion classification. By testing the proposed model on Restaurant domain reviews of the SemEval2016 benchmark dataset, experimental results showed that our proposed sentiment analysis model has achieved significantly better accuracy than the baseline model.