Aspect-level Sentiment Classification Based on Retrieval and Local Contextual Syntactic Features

Aspect-level sentiment analysis is crucial in sentiment analysis, aiming to detect nuanced emotional tendencies towards specific topics within texts. Although pre-trained language models are widely used, there is still a need to improve the capture of syntactic details. The performance of mainstream methods is limited by the selection of contextual samples, failing to fully exploit the potential of the models. This paper proposes a framework that addresses this issue through two stages of retrieval-based example sorting and instruction tuning: example retrieval and task reasoning. During the example retrieval stage, the system selects and sorts the most relevant instances from a large sample pool. Contrastive learning enhances the performance and identification of the retriever. In the task reasoning stage, the retriever provides the most matching examples, which are then fed into the LCFS-DeBERTa model for accurate sentiment classification, integrating contextual processing. This framework significantly improves the performance and generalization of sentiment classification by precisely matching and efficiently processing examples.