Chinese spatial relation extraction model by integrating geographic semantic features

Spatial relation extraction is a crucial task in geographic information science. To address challenges such as semantic ambiguity and the sparse distribution of spatial relations, this study proposes PURE-CHS-Attn, a Chinese spatial relation extraction model integrating geographic semantic features into a deep learning framework. The model enhances the baseline PURE architecture by incorporating geographic semantics, including geographic entity types and spatial relation feature words, which are fused using a novel combination of vector concatenation, weighted averaging and attention mechanism. The model leverages BERT-wwm-ext for encoding and a multi-layer perceptron for spatial relation classification, enabling improved semantic representation and relational inference. Experimental results demonstrate that PURE-CHS-Attn achieves notable improvements over baseline methods, with a 7.0% increase in precision, a 6.5% improvement in recall, and a 6.7% boost in F1 score. The model shows significant advancements in classifying spatial topological and directional relations, though challenges persist in handling sparse relation types such as distance relations. This study contributes a methodology for integrating external geographic knowledge with deep learning, advancing the automation and precision of geographic information extraction and supporting intelligent GIS and knowledge-driven applications.