Knowledge graph embedding via reasoning over entities, relations, and text

Knowledge graph embedding has attracted significant research interest in the field of intelligent web, which aims to embed both entities and relations into a low-dimensional space. In particular, there are two fundamentally different kinds of models, latent feature models and graph feature models, to infer new predictions in the graph. Latent feature models are expert at using latent features of entities to explain triples and infer these features automatically from the data, while graph feature models are do well in extracting features from the observable graph patterns. Combining the strengths of these two fundamental models is a promising approach to increase the predictive performance of graph models. Thus, we propose a new combined model, named as Text-enhanced Knowledge Graph Embedding (TKGE), to perform inference over entities, relations, and text. The model is not only well-suited for modeling interactions of their latent features, but also well-suited for modeling paths between entities in the graph. Experimental results show that TKGE has significant improvements compared to baselines on two tasks: knowledge graph completion and triple classification.