Semantic-enhanced relation modeling for fine-grained aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) is a fine-grained task that aims to identify the sentiment polarity associated with specific aspects in a sentence. To address ABSA tasks, state-of-the-art methods incorporate attention mechanisms with Abstract Meaning Representation (AMR) to model the relationships between specific aspects and their contextual information. However, existing ABSA datasets are limited in scale and difficult to acquire. Furthermore, extracting semantic information from AMR remains challenging due to its inherent structural complexity, and the performance of ABSA is highly dependent on the quality of AMR parsing. To overcome these challenges, we propose a novel Semantic Enhanced Relation Aggregation Network (SERAN), which encodes AMR to extract semantic information and integrates global features to enrich semantic representation. First, the Word Replacement (WR) module generates adversarial samples by randomly replacing words in sentences, thereby augmenting the dataset and enhancing its lexical diversity. Furthermore, the semantic iteration module employs a self-attention mechanism to compute attention score matrix of the sentence, which is subsequently encoded by using graph convolution to enhance the semantic representation in AMR. Additionally, a relation aggregation module is designed to further enhance global sentence features by leveraging semantic information derived from AMR. Finally, extensive experiments demonstrate that the proposed model outperforms several state-of-the-art methods on four benchmark datasets.