Abstract: Aspect sentiment triple extraction aims to extract aspects and their corresponding opinion words and sentiment polarities in the form of triples from sentences.Existing extraction models suffer from issues such as insufficient exploitation of syntactic and semantic information in sentences and incorrect identification of multi-word entity boundaries.To address these issues,this paper proposes a span extraction model that deeply integrates syntactic and semantic features(Span Extractor Incorporating Semantic and Syntax Features,SESS).SESS combines self-attention mechanisms with multi-channel graph convolutional networks to deeply explore the associations between syntactic and semantic features,enhancing the model’s ability to handle complex sentence structures and multi-word entities.Additionally,the model employs a span-based extraction method to extract aspect and opinion words,capturing the overall semantics of long entities and reducing sentiment inconsistency issues.The experiments conducted on the standard dataset ASTE-Data-V2 demonstrate that SESS outperforms the vast majority of comparison models in terms of F1 score,particularly in processing complex sentences and one-to-many,many-to-one sentiment relationships.Furthermore,ablation experiments and case analysis validate the effectiveness of each module of the model and its contribution to task performance,further proving the advancement and robustness of the proposed method.

Key words: Aspect sentiment triplet extraction, Graph convolutional network, Self-attention mechanism, Dependency syntactic relationship