Enhancing aspect category sentiment analysis via prompt-based prediction with semantic augmentation

Abstract Aspect category sentiment analysis (ACSA) involves identifying sentiment polarities of aspect categories explicitly or implicitly mentioned in sentences. However, existing methodologies continue to face challenges of data scarcity and insufficient contextual understanding in short text sentiment analysis, with notable accuracy decline when processing implicit aspect expressions. This paper proposes a prompt-based RoBERTa-semantic model (PRSM) to address these complex issues. PRSM transforms the ACSA task into a masked prediction problem through customized prompt learning templates, then leverages RoBERTa to extract contextual features and predict aspect categories. The introduction of a BiLSTM-based feature fusion structure effectively alleviates the problem of insufficient semantic representation, thereby enhancing the capture of multi-granular semantic features and deep semantic dependencies. Furthermore, a T5 autoregressive model generates explanatory text, compensating for inadequate understanding of implicit information. Experimental results across four benchmark datasets demonstrate that the proposed model significantly outperforms baseline approaches.