Ontology Embeddings for Subsumption Prediction Based on Graph Language Model

With the growing importance of knowledge graphs in artificial intelligence, accurately modeling hierarchical relationships in ontologies has become a critical issue in knowledge representation learning. To address this, this paper proposes an ontology embedding framework based on graph language models, named GLMSubs, aimed at enhancing the prediction of subclass relationships. The GLMSubs framework adopts a two-stage strategy of “multi-semantic view partitioning” and “advanced training of graph language models”. Initially, it deconstructs the ontology's concepts, attributes, and instance information into five types of semantic views, such as class hierarchy view and class-attribute relationship view, through a multi-view partitioning mechanism, comprehensively capturing information from different semantic dimensions. Subsequently, the framework employs graph language models for joint training on the multi-view data to obtain embeddings that integrate both semantic and structural information. Experiments on datasets such as FoodOn and GO validate the effectiveness of GLMSubs, demonstrating that its performance in class hierarchy relationship prediction tasks significantly surpasses existing methods.