Research On Knowledge Graph Representatic Earning Incorporating Entity Description Based On DeLearning

Knowledge graph is a large-scale semantic network which represents knowledge in the form of triples like(entity,relationship,entity).It has been widely used in information retrieval,question answering over knowledge graph,and personalized recommendation system.Knowledge graph representation learning aims to represent entities and relations as low-dimensional real-valued vectors.This research can effectively alleviate the data sparseness of large-scale knowledge graph,reduce the computational complexity,so it be conducive to knowledge graph completion,knowledge inference and other tasks.At present,the typical model of knowledge graph representation learning is translating model represented by TransE.By learning the triple structure information,entities and relations can be represented vectors.Simultaneously,Each entity in the knowledge graph corresponds to a paragraph of description to describe specific meaning of the entity,which includes rich semantic information.As a complement to the triple structural information,entity description can effectively improve the performance of knowledge graph representation learning.However,most work of knowledge graph representation learning incorporating entity description,like DKRL and other models,often use word-bag models,etc.which fail to capture contextual information.In addition,these work do not relate to relations of triples when learn entity representation base on description.Actually entities often have different semantics under different relations,so entity representation should distinguish different relations.To solve the above problems,the paper designs knowledge graph representation learning models based on translation model that incorporate entity descriptions based on deep learning,as follows:(1)Knowledge graph representation learning model incorporating entity descriptions based on convolutional recurrent neural network(T-CRNN).Based on TransE,T-CRNN uses convolutional recurrent neural network to learn entity representation based on description,then incorporates entity representation based on description into knowledge graph representation learning,and get vectors representing entities and relations by training.In CRNN,the inputs are word vectors corresponding to each word in entity description.A CNN is used to extract the features in the entity description and pooling operation is used to retain the effective features and reduce the number of parameters.Then a RNN is used to get the contextual vector of entity description.(2)Knowledge graph representation learning model incorporating entity description based on attention mechanism(T-CRAN).Based on the architecture of T-CRNN,T-CRAN adds attention mechanism into convolutional recurrent neural network and associates entity representations based on description and relations of triples.Then it uses the same method as T-CRNN to train the model and obtain vector representations of entities and relations.For a specific relation,attention mechanism will pay more attention to the relationship-related features of the entity description and assign a greater weight to it.The paper adopts gradient descent algorithm to train models on two datasets,FB15K and WN18,then uses link prediction and triple classification to evaluate the representation of entities and relations.The experimental results show that the performance of T-CRNN and T-CRAN model in three indicators of mean rank,hit@10 and the classification accuracy is better than TransE model and DKRL model.