Extracting individual trajectories from text by fusing large language models with diverse knowledge

Abstract

Individual trajectories offer insights into human mobility, with data either passively recorded, such as GPS, or actively recorded, such as natural language text. While the former provides detailed movement data, it lacks important context such as personal experiences, which can be obtained from the latter. Extracting trajectories from text can enhance travel experience optimization, historical analysis, and pandemic management. However, existing trajectory extraction methods rely on rule-based frameworks that fail to capture contextual semantics, resulting in limited generalizability and loss of trajectory semantics. While general-purpose large language models (LLMs) demonstrate potential for contextual reasoning capabilities, their deficient domain-specific knowledge pertinent to trajectory patterns hinders efficient and precise trajectory extraction. To address these limitations, we propose T2TrajLLM, a novel framework that fuses LLMs with domain knowledge through three components: (1) a lightweight trajectory model for structured guidance, (2) a text-to-trajectory transformation model enabling multi-step reasoning, and (3) labelled text-trajectory samples for learning domain-adaptive constraint rules. Central to T2TrajLLM is a prompt method that dynamically fuses these components with LLMs while avoids rigid rule dependency. Evaluated across three heterogeneous datasets, T2TrajLLM achieves ∼8 % higher accuracy than existing methods, demonstrating strong transferability across datasets and extensibility to diverse application requirements. Overall, T2TrajLLM effectively extracts trajectories from diverse textual sources, providing robust support for the analysis and understanding of individual mobility.