Hypergraph-based analysis of toponym co-occurrence in web documents

Toponym co-occurrence—the appearance of multiple place names within a narrative unit—implies latent spatial, economic, or administrative connections. Extracting and modeling these co-occurrences helps decode implicit geographical knowledge from unstructured text. However, traditional network models reduce multi-way interactions to pairwise edges, thereby obscuring event-level dependencies embedded in narratives. To address this limitation, this paper establishes an event-preserving hypergraph framework for modeling toponym co-occurrence. Unlike ordinary graph-based approaches, hyperedges can simultaneously connect multiple co-occurring toponyms, thereby preserving the structural configuration of multi-way geographic events. Using the 2023 People’s Daily online news corpus, we constructed weighted toponym co-occurrence hypergraphs and investigated their higher-order centrality, scale-free heterogeneity, and small-world architecture. Furthermore, we introduce domain-specific spatial mining methodologies, including node-to-hyperedge random walks (HG-PageRank) and hyperedge-level distance decay analysis, to characterize geographical co-occurrence patterns. Under a common graph convolutional network (GCN) training and evaluation framework, incidence-based hypergraph embeddings outperform pairwise-projection embeddings in terms of topological separability and in a classification task using frozen embeddings to distinguish observed toponym co-occurrence sets from constructed alternatives. Our findings demonstrate that this modeling paradigm effectively captures the higher-order complexity of geographic associations, providing an event-preserving approach to revealing toponym co-occurrence patterns and spatial attention in web documents.