Entity Description With Relation Path Modeling For Knowledge Graph Completion

Knowledge graphs are used to describe the entities and the relations between entities in the real world.The original purpose of the knowledge graph is to improve the search quality and search efficiency.They are widely used in intelligent search,intelligent question and answer,and other fields.The knowledge graphs contain a large number of fact triples consisting of entities and relations.However,due to the sparsity of data in large-scale knowledge graphs,knowledge in knowledge graphs is incomplete,and many implicit knowledge has not yet been acquired.knowledge graphs requires the task of mining knowledge and completing the knowledge graphs.The initial methods of knowledge graph completion are using logic rules for knowledge reasoning.In general,some methods use the first--order Horn clause or first-order logic to predict the relationship between entities to complement the knowledge graph.However,inference rules based on logic rules need to be written by humans and the efficiency is low.The most commonly used methods of complementing the knowledge graphs are based on the translating embedding.The main idea of the translating embedding is to embed the entities and relations into a continuous low-dimensional vector space.The methods based on translating embedding have simple models and which are easy to operate,especially in sparse knowledge graphs.However,the translating embedding methods rely on pure data drive,and the accuracy of the prediction results are limited by the data.There are more commonly used complementary algorithms based on relation path methods.According to the characteristics of the directed graph of the knowledge graphs,this method use the random walk algorithm to calculate the probability of the relations between the entities.And completing the knowledge graphs by ranking the probabilities.The method based on the relational path can explain the prediction results well,and have better performance than the method based on translating embedding.Despite this,the method based on relation paths couldn't show good performance on sparse datasets.Especially the extraction of relation paths and the computations is expensive.In this paper,we propose a new algorithm for knowledge graph completion.First of all,our method take advantage of the textual description information in the knowledge graphs.We combine the word vectors in the text and use the text vectors to represent the entities.Secondly,our method make use of the relation paths information between entities in the knowledge graph.More specifically,we apply the path information in to enhance the representation of the relations between the entities.Similarity measure function of the knowledge graph's triples is derived from combining the text description information and the relation paths information.The algorithm uses a margin-based loss function to segment the positive and negative cases in the dataset for optimal calculation.And we using the batch random gradient descent method to train our algorithm.After that,we get the vector representation of entities and relations.Finally,we actualize the semantic calculation of knowledge by these vector representation to complete the task of links prediction in knowledge graph.Our method makes full use of the characteristics of the knowledge graph.We combine the relation paths information between a large number of entities and the text description information of the entities in the knowledge graph.These valuable information are applied to embed the entities and relations into a continuously low-dimensional vectors space.Vector representation of entities and relations can do the task of complementing the knowledge graph effectively.In the end,our algorithm compares and analyzes the experimental results on the standard experimental datasets through the evaluation criteria such as entity prediction and relation prediction.Through experimental evaluations,it proves that our algorithm is feasible and effective.