Multi-task specialized expert model for hierarchical aspect-based sentiment analysis in consumer healthcare

Patient reviews offer valuable insights into patients’ experiences and perceptions of healthcare services. However, sentiment analysis of these reviews remains challenging due to limited classification granularity, high annotation demands, and the suboptimal performance of existing Aspect-Based Sentiment Analysis (ABSA) methods. These constraints hinder the precise identification of clinical departments in need of improvement and limit the fine-grained understanding of patients’ dissatisfaction and expectations, thereby restricting targeted healthcare management. To overcome these challenges, we propose a Hierarchical Aspect-Based Sentiment Analysis (H-ABSA) framework tailored to healthcare contexts. H-ABSA introduces a clinically-informed hierarchical labeling system comprising five primary aspects and twelve secondary sub-aspects, enabling multi-level sentiment analysis from both semantic and clinical perspectives. Furthermore, we develop a Multi-task Specialized Expert Model (MuSEM) that coordinates multiple Expert Models (EM) via a training-free directed routing mechanism. To enhance cross-task semantic understanding and feature sharing, MuSEM integrates Dual and Triple Cross-Attention (DCA and TCA) modules. Experimental results demonstrate that our approach notably improves both model performance and sentiment granularity. It achieves micro and macro F1 scores of 0.82 and 0.81, respectively, in end-to-end (E2E) ABSA tasks. This work establishes a practical and extensible framework for consumer-oriented feedback analysis, offering meaningful implications for data-driven healthcare management.