Semantic Relatedness Enhanced Graph Network for aspect category sentiment analysis

As a variant problem of aspect-based sentiment analysis (ABSA), aspect category sentiment analysis (ACSA) aims to identify the aspect categories discussed in sentences and predict their sentiment polarities. However, most aspect-based sentiment analysis (ABSA) research focuses on predicting the sentiment polarities of given aspect categories or aspect terms explicitly discussed in sentences. In contrast, aspect categories are often discussed implicitly. Additionally, most of the research does not consider the relations between contextual words and aspect categories. This paper proposes a novel Semantic Relatedness-enhanced Graph Network (SRGN) model which integrates the semantic relatedness information through an Edge-gated Graph Convolutional Network (EGCN). We introduce an ontology-based approach and a distributional approach to calculate the semantic relatedness values between contextual words and aspect categories. EGCN with the capability to aggregate multi-channel edge features, is then applied to model the semantic relatedness values in a graphical structure. We also employ an aspect-context attention module to generate aspect-specific representations. The proposed SRGN is evaluated on five datasets constructed based on SemEval 2015, SemEval 2016 and MAMC-ACSA datasets. Experimental results indicate that our proposed model outperforms the baseline models in both accuracy and F1 score.

Automated summary

This paper proposes a novel approach, SRGN, that integrates semantic relatedness information into aspect category sentiment analysis. By modeling the semantic relatedness between contextual words and aspect categories using the EGCN model and calculating relatedness values using ontology and distributional approaches, SRGN outperforms baseline models in terms of accuracy and F1 score on multiple datasets.