Ontology-Constrained LLM Scaffolding for Task Initiation

Sammanfattning

Task-initiation difficulty occurs when people intend to act but struggle to begin. While just-in-time adaptive interventions (JITAIs) can deliver support at relevant moments, most rely on predefined message libraries that limit personalization. Recent work has proposed replacing these with large language model (LLM) generations but delegating both selection and generation to the model risks making the decision logic less explicit. This thesis examined whether task-initiation difficulty can be represented in an ontology that constrains LLM scaffold selection before generation. A conventional inductive content analysis of 11 Reddit threads identified task-initiation blockers and scaffolds, which were grouped into 12 categories, filtered for brief text-based delivery, and formalized into an ontology grounded in an epistemic compatibility framework. The ontology was implemented in an app-based, JITAI-inspired multi-agent system. A one-week pilot study with five adult participants (two women, three men; ages 27–62) evaluated feasibility through the Technology Acceptance Model (TAM) questionnaire, interviews, and system-use logs. Participants found the system relatively easy to use, while perceived usefulness was mixed. The feedback loop functioned end-to-end but produced too little data for meaningful personalization within the study period. The findings suggest feasibility rather than effectiveness. More broadly, this thesis illustrates how linking represented blocker states to ontology-constrained scaffold options before LLM generation can make adaptive intervention selection more explicit and controllable, which may be especially relevant in health-related settings.