Enhancing consumer health question answering systems through rhetorical structure theory-guided large language model

Abstract

Consumer Health Question Answering (CHQA) systems have traditionally been designed around isolated capabilities—either factual medical correctness or empathetic support—rather than delivering integrated, multi-dimensional assistance. This separation fails to reflect the reality of consumer health consultations, where individuals typically express intertwined medical concerns, personal circumstances, and emotional distress within a single query. Existing CHQA architectures, often optimised for short, single-intent inputs, prioritise answer generation over question understanding and lack explicit mechanisms to jointly interpret informational and emotional support needs. Although Large Language Models (LLMs) have advanced the state of medical QA, most LLM-based systems still struggle to reconcile clinical reliability with context-sensitive empathy in complex consumer health questions.

To address this gap, this study adopts a Design Science Research (DSR) paradigm and introduces Joint Medical–Emotional Question Answering (JMEQA) as a new task that jointly understand and response medical informational needs and emotional support needs in Consumer Health Questions (CHQs). Grounded in Rhetorical Structure Theory (RST) and Appraisal Framework for Clinical Empathy (AFCE), the research proposes the Med-Emo CHQA architecture, a neural-symbolic unified, LLM-based system that augments answer generation with explicit question understanding. The architecture integrates hierarchical modules for question understanding and answer generation, unifying symbolic knowledge (an RST-based discourse structure tree that encodes intents, contexts, and rhetorical relations during the question-understanding stage) with neural network (LLM) via supervised fine-tuning and structured prompting.

To support system training and evaluation, this study constructs CHQA-MedEmo, the first large-scale, multi-layer corpus of Chinese health consultation records. The corpus is annotated for both medical informational needs and emotional support needs, with explicit contextual spans and discourse relations. System performance is assessed through controlled experiments and an expert reader study using a human-centred evaluation framework that jointly measures accuracy, personalisation, and empathy, alongside latency. Experimental and reader-study results show that the proposed architecture outperforms baseline and ablated models in full medical-need coverage and empathetic alignment, while maintaining lower latency.

Overall, this study advances the design of CHQA systems by demonstrating how symbolic, discourse-aware question understanding can be embedded into LLM workflows to produce responses that are simultaneously medically reliable, personally tailored, and emotionally supportive. The findings contribute design knowledge for Neural-Symbolic Integration (NSI), LLM-based CHQA in online health consultation settings and offer a scalable architectural paradigm for building complex, human-centred AI systems that balance informational rigour with affective care.