Where Guidance Happens: Toward a Knowledge Graph of Situated Orientation and Mobility Instruction

Abstract

Orientation and mobility (O&M) training helps blind and low-vision people travel safely and independently, but one-on-one specialist instruction is difficult to scale. Wearable sensors could extend support into independent practice. However, models for real-time instruction require structured examples linking mobility events and environmental context to specialist decisions. We address this gap using interviews with 10 blind and low-vision adults and 10 O&M specialists. A human-in-the-loop workflow uses a locally deployed large language model to extract candidate concepts from interview transcripts, assign them to researcher-defined semantic categories, and propose candidate relationships, followed by grounding checks, semantic verification, schema normalization, and researcher review. The resulting provenance-preserving knowledge graph represents learner state, mobility activity, environment, hazards, sensory cues, personalization factors, and instructional strategies while retaining links to supporting quotations. Competency-driven SPARQL queries evaluate what the graph can represent and retrieve across instructional and contextual dimensions. This work provides a semantic foundation for future event-level data collection and specialist-supervised predictive models for O&M instruction.