Span Pruning and Syntactic Awareness for Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) is a composite task aimed at extracting key information from sentences and determining sentiment polarity. While existing methods have shown promising results, many fail to utilize the relationships between span pairs and lack effective pruning strategies for extracting aspect and opinion terms. In this paper, we propose SPASA, a model that combines span pruning and syntax awareness, divided into two stages: Named Entity Recognition (NER) and Relation Extraction (RE). In the NER stage, levitated markers, Bi-affine layers, and attentive pooling layers are used to model boundaries and semantics of the entity spans, with BCEWithLogitsLoss filtering low-quality spans. In the RE stage, a Graph Convolutional Network (GCN) is introduced to capture syntactic dependencies and enhance sentiment triplet extraction by modeling subject-object span pair interactions. Experimental results demonstrate that SPASA outperforms baseline models in recall rate and F1 score across four standard datasets, with improvements of 1.79% and 4.67%, 1.95% and 4.27%, 1.7% and 3.13%, and 1.78% and 1.75%, respectively, highlighting the effectiveness of span interactions in sentiment triplet extraction.