부산대학교 도서관

Aspect Sentiment Triplet Extraction (ASTE) aims to extract aspect terms, corresponding opinion terms, and the expressed sentiment polarities from a given sentence. Recent endeavors have explored the utilization of token-level semantic interactions, yielding enhanced performance in the ASTE task. However, these approaches dissect the semantic information of entire spans without fully capturing the contextual interaction of aspect/opinion terms, thus lacking the ability to handle complex linguistic structures, encompassing multi-word terms and sentences featuring multiple triplets. In this paper, we present a novel Span-level Contextual Interaction Network (SCI-Net) to efficiently model bi-directional contextual interactions between aspect and opinion terms. Specifically, SCI-Net first employs linear projection layers to meticulously construct discrete, task-oriented token representations for aspects, opinions, and sentence-level contextual semantics, respectively. Complementing this, we integrate a cross-attention mechanism, with the contextual semantics serving as a dynamic mediator, dynamically fusing contextual semantics with aspect and opinion representations, thereby facilitating bi-directional contextual interactions. Additionally, we leverage local context between aspect–opinion pairs to further refine sentiment polarity prediction. Experiments on four benchmarks prove that our model yields state-of-the-art performances.