Aspect Sentiment Triplet Extraction via Affective Knowledge and Enhanced Graph Attention Networks

Abstract

Aspect Sentiment Triplet Extraction (ASTE) refers to a new and emerging task for sentiment analysis and it is designed to extract triplets consisting of aspect terms, opinion terms, and sentiments from sentences that are user comments. While Graph Convolution Networks (GCN) and external emotional knowledge have shown promise, they have limitations in assigning node weights and capturing intricate relationships. Particularly, GCN treats neighboring nodes uniformly during convolution, lacking the capability to allocate distinct weights based on node significance. In this paper, we propose a SenticNet enhanced Graph Attention neTworks (SN-GAT) model. Specifically, we employ Graph Attention neTworks (GAT) by integrating node and edge information to thoroughly account for variations in the importance of different nodes, acknowledging that certain nodes hold a more significant influence on the overall structure or outcome compared to others. Additionally, our model employs external affective knowledge from SenticNet, a resource encompassing emotions, sentiments, and affective associations in language, to refine word pair representations. This strategy enhances implicit results in aspect and opinion extraction. By conducting extensive experiments on baseline datasets, our proposed model demonstrates superior performance compared to other methods.