Synthesizing Geographical Semantics and Large Language Model for OD ‐Based Trajectory Captioning

Trajectory captioning aims to represent trajectories in natural language texts, providing a human‐friendly approach to understanding and utilizing trajectory data. Previous trajectory captioning methods relied on manually defined rule templates to describe trajectories, failing to accurately and comprehensively capture the trajectory semantics. To address these issues, this paper synthesizes the semantics of the origin and destination (OD) points of trajectories with large language models (LLMs) to improve trajectory captioning. The proposed method first designs a trajectory context augmentation module that enriches the trajectories' context using the spatial information from Point‐of‐Interest (POI) distributions and human travel data from check‐in records. The method develops a geographical semantic‐based reasoning module that constructs a time‐subcategory graph and trains a POI matching module to capture the geographical semantics of the trajectories. Finally, a trajectory decoding module is developed, which leverages LLMs and the Chain‐of‐Thought inspired strategies to generate trajectory captions. Extensive experiments show that the proposed method outperforms baseline methods on both lexical‐overlap and semantic‐alignment metrics. This paper explores a new approach for trajectory semantics representation and provides a methodological reference for the text generation tasks in various geographical applications.