Large-scale Knowledge Representation Learning Based On Sub-relation Path Features

As an important technology of knowledge graph completion,knowledge reasoning has become a hotspot in current research.The path-based Knowledge Representation Learning(KRL)takes into account both the semantic information of the relation paths and the entities,which greatly improves the accuracy of knowledge reasoning.It models complex relation path to improve the accuracy of knowledge reasoning,thus increases the time overhead and is difficult to adapt to the increasingly large-scale knowledge graphs.Aiming at the efficiency problem of complex relation path modeling,a method for modeling Sub-relation Path Features(SPF)is proposed.The main idea of SPF is to use the Length-limited Breadth-First Search(LBFS)algorithm to extract sub-relation path features for each entity in the knowledge graph and form a sub-relation path feature set.It is only necessary to selectively combine sub-relation path features according to the features of neighbor entities,instead of modeling the full relation path directly,this improves the efficiency of relation path modeling.SPF can decrease the complexity from exponential level to constant level.In addition,SPF can reduce redundant computation of the complex relation path modeling.Although SPF can effectively decrease the time overhead,when the relation path length is long,SPF needs to record more sub-relation path features,which causes large memory overhead.Therefore,for long relation paths,a limited number of paths method with SPF(LSPF)is proposed.LSPF can decrease the space overhead by filtering out relation paths that are not helpful for knowledge reasoning.Experiments show that SPF can reduce the time overhead by 27.1% to 59.5% while ensuring the accuracy of knowledge reasoning compared to complex relation path modeling;when modeling longer relation path,LSPF can reduce the space overhead of SPF by 1.65 times,and reduce the time overhead of complex relation path modeling by 69.8%.