STMC-GCN: A Span Tagging Multi-channel Graph Convolutional Network for Aspect Sentiment Triplet Extraction

Abstract

Aspect-Based Sentiment Triplet Extraction (ASTE) is a rapidly growing field in sentiment analysis. While most research has focused on processing the ASTE task either in a pipeline or end-to-end manner, both methods have their limitations. Pipeline methods may accumulate errors in practical applications, while sequence labeling methods in end-to-end approaches may overlook important feature information of the three elements themselves. Additionally, various features in sentences and emotional word markers have not been effectively explored in these methods. To address these limitations, we propose a novel solution called Span Tagging Multi-Channel Graph Convolutional Network (STMC-GCN) that explicitly combines multiple prominent features to extract span-level sentiment triplets, where each span may consist of multiple words and play different roles. Specifically, we designed a three-channel graph fusion model that converts sentences into multiple channels of graphs. These channels extract node text features, centrality features, and position features, which are then extracted through cross-channel convolution operations to obtain a common graph representation shared by different channels. To optimize downstream classification with better results, we use consistency and difference constraints to enhance common attributes and independence. Finally, we explore span-level information and constraints to generate more accurate aspect-based sentiment triplet extractions. Experimental results illustrate that STMC-GCN performs well on multiple datasets, proving the effectiveness and robustness of the model.