Multi-granularity features fusion with hierarchical networks for aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) aims to identify and classify sentiment polarities toward specific aspects in text. To address the limitations of existing models, such as oversimplified structures and underutilized fine-grained features, we propose an ABSA model that fuses multi-granularity features with hierarchical networks. Through the construction of a multi-granularity feature extraction module, we employed a graph convolutional network (GCN) on constituent and dependency trees to capture syntactic and semantic information, augmented with external knowledge. This enables comprehensive feature extraction from four granularity levels: constituent structure, dependency relations, contextual cues, and external knowledge. To effectively fuse these diverse features, we designed a multi-layer feature fusion network (MLFF). Utilizing cross attention and orthogonal projection, the MLFF module iteratively refines feature interactions. Extensive experiments on the SemEval 2014 and Twitter datasets show that our model outperforms existing ABSA models and can effectively enhance task performance.