A unified dual-view knowledge-guided sentiment interaction networks for aspect-based sentiment analysis.

Abstract

Aspect-based sentiment analysis enables precise identification of sentiment-bearing entities and attributes in textual content, thereby delivering granular sentiment information for downstream applications. Recent research has focused on enhancing sentiment-syntactic dependencies through external knowledge and graph neural networks. However, effective augmentation of sequence semantics via external knowledge remains underexplored, and existing methods inadequately capture the complementary relationship between sentiment semantics and syntax. To address these limitations, we propose the Dual-view Knowledge Guided Sentiment Interaction Network (Dual-KGIN). The core contribution of Dual-KGIN lies in the construction of a unified framework for enhancing both sentiment semantics and syntactic representations by integrating external knowledge and implementing hierarchical interactions. Specifically, Dual-KGIN designs an external knowledge-guided syntactic GCN module that uses external knowledge to refine adjacency dependencies and enhance syntactic feature learning. Secondly, we use external knowledge based on the original attention mechanism to augment sequence semantics. Finally, we propose a novel multi-level feature interaction module, which effectively enhances the interaction of different perspectives at multiple levels of sentiment knowledge to improve sentiment representation. Extensive experiments on benchmark ABSA datasets demonstrate Dual-KGIN's superior aspect-specific sentiment identification capability. Ablation studies validate the efficacy of each component, highlighting Dual-KGIN's ability to effectively harness knowledge and model feature interactions for state-of-the-art ABSA performance.