Scholarlink

This study presents a sentiment-analysis framework that integrates a large language model with a structured sentiment knowledge graph to evaluate smart-home product reviews. Moving beyond the conventional three-class scheme of positive, negative, and neutral sentiment, the framework adopts an eight-emotion taxonomy that captures fine-grained affective variation. Product features and their associated emotions are automatically extracted from reviews and encoded as nodes and edges in a graph, whose prior information is subsequently injected into the language model through prompt optimization and adaptive weighting. On a manually annotated test set the proposed method attains an accuracy of 82.05%, exceeding the performance of all baseline models. The results demonstrate that coupling graph-based priors with a modern language model improves the detection of domain-specific sentiments, providing consumers with precise insight into product perception and offering manufacturers reliable evidence for product refinement.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SKG-LLM: Enhancing Large Language Models with Sentiment Knowledge Graphs for Fine-Grained Sentiment Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yixuan Yuan,
  • Bixuan Li

摘要

This study presents a sentiment-analysis framework that integrates a large language model with a structured sentiment knowledge graph to evaluate smart-home product reviews. Moving beyond the conventional three-class scheme of positive, negative, and neutral sentiment, the framework adopts an eight-emotion taxonomy that captures fine-grained affective variation. Product features and their associated emotions are automatically extracted from reviews and encoded as nodes and edges in a graph, whose prior information is subsequently injected into the language model through prompt optimization and adaptive weighting. On a manually annotated test set the proposed method attains an accuracy of 82.05%, exceeding the performance of all baseline models. The results demonstrate that coupling graph-based priors with a modern language model improves the detection of domain-specific sentiments, providing consumers with precise insight into product perception and offering manufacturers reliable evidence for product refinement.