A Study On Text-augmented Knowledge Representation Algorithm Based On Gated Convolutional Network

The purpose of knowledge graph representation learning is to project entities and relations into low-dimensional continuous vector space,so the knowledge graph is compatible with machine learning model.Knowledge graph completion is the task of predicting the missing relations between entities.There are a lot of important text information about entity description in the knowledge base,while the existing models only consider the knowledge triples indicating the relations between entities,but does not consider the descriptions of entities.Therefore,this paper proposes a text-augmented model based on gated convolutional neural network(GConvTA).In the model,the structure representation of triples and the text representation encoded by entities descriptions are fused,which can learn from symbol triples and text descriptions together,and can establish the join and interaction between fact triples and text descriptions.All triples(head entity,relation,tail entity)are represented as 3-column structure embedding and 3-column text embeddings obtained from the descriptions of memory network with attention by pretraining.On the one hand,text representation and structure representation are integrated in the input layer,then the fused feature map is obtained.On the other hand,feature interaction can also be carried out in the gated convolution layer,that is,structure embedding and text embedding can be input into convolution respectively,and in the gated unit,through point multiplication operation,the two represented feature maps can be interactive.Finally,the two-stage feature maps are fused again to get the vector which represents the triple.This feature vector gets the final score through point multiplication operation to judge whether the triple is valid or not.The link prediction results show that GConvTA model is better than previous baseline models such as ConvKB,ConvE on FB15k-237 and WN18 RR benchmark data sets.In zero-shot,each index is higher than DKRL model on FB20k,which further shows the superiority in text-enhanced knowledge graph model.