Understanding Navigation‐Lost Scenes Based on Knowledge Graphs Using a Deep Learning Approach

ABSTRACT Disorientation is a common and impactful failure mode in everyday navigation. This paper addresses this issue by modeling Navigation‐Lost (NL) scenes as structured interactions between human and environmental factors. We propose the Navigation‐Lost Knowledge Graph (NLKG), a conceptual framework that incorporates 24 entity types and 27 relation types, covering subjective perception, travel modes, and objective environmental context (e.g., weather, terrain, road layout). To efficiently populate the NLKG at scale, we develop a triple extraction model— NavLTR —which includes modules for handling nested semantics, hierarchical relations, and urban functional categories. To mitigate data scarcity and anchor evaluation, we create the Navigation‐Lost Information Extraction Database (NIED), a comprehensive domain corpus. On extensive benchmarks, NavLTR achieves state‐of‐the‐art performance, with an F1 score of 92.82%, significantly outperforming existing baselines. We demonstrate the practical utility of NLKG in two downstream applications: (i) a prototype question‐answering system that surfaces contextual cues to assist navigators and (ii) multi‐perspective graph analytics for risk prediction and factor attribution, providing actionable insights for urban wayfinding and spatial design. Collectively, these contributions—the schema, corpus, and schema‐aware extractor—lay the foundation for a reproducible framework for NL scene understanding and offer a practical path for transforming unstructured narratives into context‐aware geospatial reasoning, benefiting both end users and urban planners.