Scholarlink

<p>Aspect-based sentiment analysis (ABSA) effectively enables fine-grained understanding of medical feedback texts, but existing methods typically lack transparency, limiting trustworthiness in practical medical scenarios. To address this issue, we propose a two-stage ABSA framework for healthcare, integrating large language models with interpretable classifiers. In the first stage, structured aspect-sentiment pairs are extracted from patient feedback using a large language model guided by chained prompts. The second stage employs a lightweight interpretable classifier to perform final sentiment predictions on these structured representations, generating explicit decision justifications through post-hoc explanation methods such as SHAP and LIME. Experiments conducted on publicly available medical datasets, including the PatientInfo patient feedback dataset, HCAHPS patient surveys, and WebMD drug reviews, demonstrate that the proposed method significantly outperforms several mainstream baseline methods. Ablation analyses further confirm the effectiveness of the prompt chaining strategy and interpretable classification. The proposed framework ensures high predictive accuracy while substantially enhancing transparency and interpretability, making it highly suitable for medical feedback analysis tasks that require model reliability and trustworthiness.</p>

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

Combining large language models with interpretable models for explainable aspect-based sentiment analysis in the medical domain

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月18日

查看全文
  • Yang Zhang,
  • Song Wen,
  • Yanping Zhu,
  • Zhijun Li,
  • Xiaohui Wang

摘要

Aspect-based sentiment analysis (ABSA) effectively enables fine-grained understanding of medical feedback texts, but existing methods typically lack transparency, limiting trustworthiness in practical medical scenarios. To address this issue, we propose a two-stage ABSA framework for healthcare, integrating large language models with interpretable classifiers. In the first stage, structured aspect-sentiment pairs are extracted from patient feedback using a large language model guided by chained prompts. The second stage employs a lightweight interpretable classifier to perform final sentiment predictions on these structured representations, generating explicit decision justifications through post-hoc explanation methods such as SHAP and LIME. Experiments conducted on publicly available medical datasets, including the PatientInfo patient feedback dataset, HCAHPS patient surveys, and WebMD drug reviews, demonstrate that the proposed method significantly outperforms several mainstream baseline methods. Ablation analyses further confirm the effectiveness of the prompt chaining strategy and interpretable classification. The proposed framework ensures high predictive accuracy while substantially enhancing transparency and interpretability, making it highly suitable for medical feedback analysis tasks that require model reliability and trustworthiness.