SDGCN: Span Dual-Channel Graph Convolutional Networks for Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) iden-tifies terms of aspect, terms of opinion, and sentiments in text. Early approaches to the ASTE task, used token-level models, were prone to errors during decoding, affecting accuracy. Span-level models address this by capturing interactions between aspect and opinion spans. But encounters challenges due to sentence encoding methods that primarily rely on contextual data without utilizing syntactic and semantic information. This limitation lead to an inaccurate interpretation of complex sentiment interactions. This study proposes a span-based model that combines syntactic encoding through dependency parsing and graph convolutional networks, along with semantic encoding using multi-head attention and graph convolutional networks. Experimental results on four public datasets show the model achieves a precision of 74.9%, recall of 67.12%, and F1 score of 71.7%, providing an average improvement of 3% over span-based methods in aspect sentiment triplet extraction.