A novel adaptive marker segmentation graph convolutional network for aspect-level sentiment analysis

Abstract

Aspect-level sentiment analysis is a fine-grained sentiment classification task that aims to identify the sentiment polarity of specific aspects in online reviews. Attention mechanisms and graph convolutional networks have recently been widely used to model associations between aspects and opinion words. However, these methods face challenges in accurately modeling the alignment of aspects and exploiting multiaspect sentiment dependencies due to the limitations of dependency trees and the complexities of online reviews. In this paper, we propose a novel adaptive marker segmentation graph convolutional network (AMS-GCN) for aspect-level sentiment analysis. Specifically, the proposed AMS-GCN model enhances the information capacity of words by merging marker information from two datasets and uses an adaptive marker segmentation module to divide different marker information into separate modules. Furthermore, the model employs bi-syntax-aware and semantic auxiliary modules to obtain syntactic and semantic information. The bi-syntax-aware module combines component and dependency trees to capture comprehensive syntactic information. In contrast, the semantic auxiliary module uses an attention score matrix to capture the semantic association information of each word. Moreover, the aspect-related graph is devised to aggregate information about the sentiment of different aspects. Experiments on several benchmark datasets demonstrate that the proposed model achieves state-of-the-art results.