Senti-EGCN: An Aspect-Based Sentiment Analysis System Using Edge-Enhanced Graph Convolutional Networks
Chen Li, Junjun Zheng, Peng Ju, Yasuhiko Morimoto · 2023
Aspect-based sentiment analysis systems aim to classify sentences according to specified aspects. Most previous studies have used graph convolutional networks (GCNs) to analyze syntactic features and used word dependencies for syntactic context and aspects. However, traditional GCNs have limitations in exploring syntactic dependency graphs, such as missing information on edges in syntactic dependency trees. For accurate sentiment prediction on specific aspects, an aspect-based senti ment analysis system using edge-enhanced graph convolutional networks (Senti-EGCN) is developed. In particular, a bidirectional long short-term memory (Bi-LSTM) network is employed to extract the contextual features of sentences. Thereafter, a transformer encoder with a self-attention mechanism is used to analyze the interrelationships and global features of words in long texts. Next, the Bi-LSTM network is used to enforce the sentence structure through the word dependency tree. The dependency tree analyzes the words’ dependencies to enhance the representation of information. A bidirectional GCN (Bi-GCN) uses message passing to propagate information across the nodes in a parsed dependency tree. In addition, an aspect-specific masking technique is applied to reduce redundant information in the hidden representation by masking contextual information outside the aspect and improve the accuracy. Finally, attention scores are calculated in the last layer for sentiment classification. The experimental results demonstrated that the proposed Senti-EGCN outperforms other baseline models in most metrics on the three benchmark datasets.