MulGCN: MultiGraph Convolutional Network for Aspect-Level Sentiment Analysis

Aspect-level sentiment analysis (ALSA) is used to identify the sentiment polarities of the given aspects in a sentence. Various approaches have been proposed to improve the performance of ALSA, most recently graph convolutional networks (GCNs). Although GCN-based ALSA methods have obtained the promised results, how to effectively and simultaneously harness the semantic, syntactic structure information from the dependency tree and the contextual affective knowledge regarding the specific aspect remains a challenging research question. This research proposes a novel sentiment analysis method applied at the aspect level, called multigraph convolutional network (MulGCN), by integrating three GCNs. Unlike previous GCNs, the MulGCN model can simultaneously capture features related to three knowledge: syntax, semantics, and context by combining the dependency parser tree, affective information in SenticNet, and inter-aspect-aware technique. The research starts with a comprehensive survey of articles related to ALSA methods based on GCN to evaluate and unify the approach, thereby identifying the point where GCN has not been adequately used in ALSA methods to have a basis for proposing appropriate improvements to improve performance. Next, three knowledge-based GCNs are built to represent and extract high-level features related to syntax, semantics, and context. Then, the fusion mechanism is used to integrate the extracted features. Finally, these features are fed into a classifier consisting of convolutional layers to determine the sentiment polarity of the aspects. The MulGCN model will be experimented on three benchmark datasets. The experimental results prove the effectiveness of MulGCN model for improving the performance of ALSA including the accuracy and the $F_{1}$ score.