Knowledge Graph-Enhanced Large Language Models for Opinion Target Sentiment Classification

Abstract

This paper addresses the challenges of tight coupling between target and opinion, significant cross-sentence dependencies, and frequent implicit sentiment and rhetorical expressions in opinion-target sentiment classification. It proposes a unified modeling method integrating knowledge graphs and large language models to achieve fine-grained sentiment discrimination for target-oriented purposes. The method obtains context-relevant text representations using a large language model and constructs a target-centric subgraph by retrieving adjacent entities and relationships from the knowledge graph around candidate targets, resulting in a structured knowledge representation. Subsequently, a lightweight gating fusion mechanism adaptively integrates the two sets of information to form a target-level fusion representation, and a concise classifier outputs the predicted sentiment polarity. This framework leverages the advantages of semantic understanding and explicit relational constraints to mitigate discrimination instability caused by target disambiguation, attribute attribution, and noisy expressions, while enhancing output interpretability and maintaining computational controllability. Comparative experiments demonstrate that this method achieves leading performance across multiple evaluation metrics, validating the effectiveness of the synergy between knowledge enhancement and semantic modeling in improving the performance of opinion-target sentiment classification.