A Reflective Storytelling Agent for Older Adults: Integrating Argumentation Schemes and Argument Mining in LLM-Based Personalised Narratives

Researchers developed a reflective storytelling agent that combines large language models with knowledge graphs and argumentation theory to generate personalized narratives for older adults. Testing with 55 participants showed the system successfully identified personally relevant purposes in two-thirds of narratives, with argument-based grounding and hallucination detection significantly improving perceived consistency and clarity.

This research addresses a critical challenge in deploying large language models for vulnerable populations: ensuring generated content remains grounded, transparent, and trustworthy. The study moves beyond generic LLM applications by implementing domain-specific safeguards through knowledge graphs and argumentation mining—techniques that force the model to justify its outputs against structured user models and formal logical frameworks.

The two-phase methodology demonstrates sophisticated research design. Domain experts in phase one ensured the system reflected real-world health priorities for older adults, while phase two's evaluation of 55 participants across multiple dimensions (purpose recognition, cultural relevance, consistency) provides meaningful validation. The 67% recognition rate for personally relevant purposes represents a substantial improvement over unguided LLM outputs, though the 50% rate for argument-based purposes suggests room for refinement in explanation quality.

The correlation between hallucination-risk indicators and human perception of inconsistency validates the argument-mining approach as an inspection mechanism rather than merely a theoretical exercise. This has broader implications for AI deployment in healthcare and elder-care contexts where accuracy directly impacts user safety and trust. The finding that cultural recognizability strongly influenced adoption intent signals the importance of localization beyond simple translation.

Future applications should explore whether these argument-grounding techniques scale to other vulnerable populations or safety-critical domains. The work establishes a replicable framework combining formal verification methods with human-centered evaluation, bridging the gap between AI transparency research and practical implementation requirements.

• →Knowledge graphs and argumentation mining effectively reduce hallucinations in LLM-generated health narratives for older adults.
• →Two-thirds of generated narratives achieved personally relevant purposes, with argument-based grounding correlating strongly with perceived clarity.
• →Cultural recognizability emerged as the primary factor determining user willingness to adopt the storytelling functionality.
• →Hallucination-risk indicators computed by the system accurately predicted human perception of narrative inconsistency.
• →The framework establishes a replicable model for deploying LLMs in safety-critical domains through formal grounding and reflection mechanisms.