Knowledge-aware interaction networks for domain-adaptive end-to-end aspect-based sentiment analysis

Great efforts have been made toward leveraging unsupervised domain adaptation techniques in end-to-end aspect-based sentiment analysis (E2E ABSA), to overcome the reliance of supervised methods on substantial labeled data. However, many existing efforts were often coarse-grained, primarily emphasizing the enhancement of domain invariance, furthermore, ignored the strong correlation between the two intertwined subtasks of aspect term extraction and aspect-level sentiment classification. To address such issues, this study proposes a novel approach, namely knowledge-aware interaction networks for domain-adaptive E2E-ABSA. Specifically, we construct domain subgraphs using external knowledge. Subsequently, we retrieve “correlative words” connecting two domains in the graph and identify “prototype words” that share syntactic and semantic similarities with the target domain. These correlative and prototype words are then utilized to edit the source words, thereby enhancing their transferability. Furthermore, we adopt an end-to-end framework that jointly performs aspect term extraction and sentiment polarity classification, while incorporating an interaction layer to facilitate mutual influence between the two subtasks. The effectiveness of our proposed approach is verified through extensive experiments on four publicly available datasets. Results demonstrate the performance superiority of our framework over state-of-the-art methods.