Numerical Reasoning for Geographic Knowledge Graph Based on Relation‐Attribute Hybrid and Entity Semantic Learning

ABSTRACT Numerical reasoning constitutes a critical component in knowledge graph reasoning, involving the utilization of entities, relations, and numerical attributes (e.g., area, temperature) to deduce new facts. However, existing methods primarily rely on attribute encoders, the associations between relations and their numerical attributes, and additional loss functions to predict numerical relations. They overlook entity semantic information, which limits their ability to understand characteristics of entities and capture semantic relations between entities, particularly when reasoning about missing relations among entities in geographic knowledge graphs. To address this issue, we propose the (Dual Relation‐Aware Embedding) DRAE model. DRAE utilizes the Entity Enhancer to incorporate semantic information into the numerical reasoning process, enabling it to better understand complex characteristics of entities and their semantic relations in geographic knowledge graph, thereby acquiring more accurate numerical representations to facilitate effective numerical reasoning. To assess the model's capability to process semantic information in geographic knowledge graphs, we create a new dataset Geography, which is adapted from FB15k‐237 and specifically incorporates semantic information designed to be conducive to numerical reasoning tasks. Compared with the existing state‐of‐the‐art methods, empirical evaluations across three datasets (Credit, Spotify, Geography) demonstrate that DRAE outperforms. DRAE achieves improvements of 11.546% in Hits@1, and 12.25% in MRR.