Research On Knowledge Graph Link Prediction By Integrated Learning On Textual And Structural Information

Knowledge Graphs(KGs)are of vital importance in multiple practical scenes,including recommender system,question-answering and information retrieval.However,human-curated knowledge graphs are usually incomplete,urging for completion by automatic mechanism,i.e.link prediction.Link Prediction is a key task to complete missing factual links between entities.Previous methods,like graph embedding approaches,mainly focus on modeling the structure information for knowledge reasoning.While their work is intrinsically sensitive to graph incompleteness,leading poor performance under sparse circumstance.Recent work,e.g.KG-Bert,attempts to utilize textual information contained in KGs to alleviate incompleteness problem.While lack of structure knowledge reduces the ability of explanation and the costly inference on total entities also prevents its large-scale applications.Some previous work,like DKRL,utilizes textual information of entities and relations as an alternative to complement the missed structural information.While these studies prefer the fusion of structural and textual embeddings,which needs elaborate and fixed model design processes and achieved poor improvements.In this paper,we proposes a novel jointly learning framework from the perspective of Predictive Distributions(PDs).Specifically,given a text-related model and a structure-based model,PDs from these two models reflect different knowledge learnt from corresponding data.Motivated by Knowledge Distillation,we propose a information fusion framework based on variational EM algorithm from the perspective of PDs.To meet our requirements for fusion process,we modify the existing Comp GCN method and propose a new Bert-based link prediction model called Query KGBert.External experiments show the effectiveness of proposed models.As for fusion framework,based on modified Comp GCN and Query KGBert models,experiments results performed on three datasets demonstrate that our proposed framework can effectively promote the completion performance compared with single information-based models.Extensive studies and analysises reveal the advantages and disadvantages of our method,which is essential for decision-making on application scenarios.