Acta Geodaetica et Cartographica Sinica ›› 2025, Vol. 54 ›› Issue (11): 2052-2067.doi: 10.11947/j.AGCS.2025.20250183

• Cartography and Geoinformation • Previous Articles    

Hierarchical multi-agent collaboration for geographic event extraction and spatio-temporal parsing

Xin HU1(), Xuexi YANG1,2(), Yifan JIANG1, Xianbin WANG1, Chen DING3, Guran XIE1, Min DENG1,2,4   

  1. 1.School of Geosciences and Info-Physics, Central South University, Changsha 410083, China
    2.Hunan Geospatial Information Engineering Technology Research Center, Changsha 410018, China
    3.Beidou Research Institute, South China Normal University, Foshan 528225, China
    4.School of Geography and Environment, Jiangxi Normal University, Nanchang 330022, China
  • Received:2025-04-28 Revised:2025-10-16 Published:2025-12-15
  • Contact: Xuexi YANG E-mail:225001019@csu.edu.cn;yangxuexi@csu.edu.cn
  • About author:HU Xin (1999—), male, PhD candidate, majors in construction and application of spatio-temporal knowledge graph.　E-mail: 225001019@csu.edu.cn
  • Supported by:
    The National Key Research and Development Program of China(2022YFB3904203);The National Natural Science Foundation of China(42271485);The Natural Science Foundation of Hunan Province, China(2024JJ1009);The Frontier Interdisciplinary Project of Central South University(2023QYJC002);The Double-Thousand Plan of Jiangxi Province(jxsq2020102062);The Open Project of Hunan Geospatial Information Engineering Technology Research Center(HNGIET2024006)

RichHTML

PDF

Abstract:

Accurately extracting and interpreting geographic events from vast amounts of text is critical for understanding real-world dynamics and constructing semantically rich spatio-temporal knowledge graph (STKG). However, the diverse expressions and inherent ambiguities of spatio-temporal information within these texts present significant challenges. To address these challenges, this paper proposes a multi-agent hierarchical collaborative method that systematically applies the reasoning capabilities of large language models (LLMs) to the task of geographic event extraction and parsing in few-shot scenarios. The core of our approach is a collaborative framework consisting of a master agent and multiple specialized sub-agents. The master agent performs adaptive task decomposition and scheduling, while the specialized agents focus on dedicated sub-tasks, including geographic event extraction, temporal element parsing, spatial element localization, and spatial location reasoning. Experiments on both a baseline dataset and a spatial reasoning-enhanced dataset constructed from DUEE demonstrate the method's superior few-shot spatio-temporal parsing capabilities. Specifically, on the baseline dataset, our method attains spatial element parsing performance of F1@100 m=0.779 and temporal element parsing performance of F1time=0.856, yielding absolute improvements of 16.3% and 22.1% over the current state-of-the-art baseline, respectively. Then, Ablation studies further validate the effectiveness of the proposed collaborative framework design. Furthermore, a case study analyzing social media data on the “July 20 heavy rainstorm in Zhengzhou” event illustrates the method's effectiveness in capturing the spatio-temporal progression and enhancing the understanding of event evolution. This study introduces a novel multi-agent processing paradigm for geographic event extraction, offering robust support for the construction of refined STKG and the advancement of event-driven geospatial intelligence.

Key words: geographic event, event extraction, spatio-temporal parsing, multi-agent, large language model

CLC Number: