Research Of Knowledge Graph Completion Algorithm

Knowledge graph is an important research field of artificial intelligence.It provides service for many application,such as machine reading comprehension,question answering and information retrieval.Knowledge graph is used to store facts in the real world,which is a network structure composed of nodes and edges.Nodes represent entities and edges connected between nodes represent the relationship of entities.Although existing knowledge graphs contain a lot of facts,they are still far from completeness.The incompleteness of knowledge graphs has serious implications for downstream applications.Therefore,it is necessary to research completion algorithm for knowledge graph to infer the latent facts.There are two different ways for knowledge graph completion in our paper.One way is using external data sources to extract new knowledge.And the other way is finding the connection patterns of nodes based on the existing data of knowledge graphs.The main work and innovation of this paper are as follows:(1)When using external data sources to complete knowledge graph,we find that previous approaches treat the task of relation extraction as a multi-class classification problem.They firstly take a sentence as the input,then extract a feature vector,finally feed it into the classifier to get the final result.Actually,a relation type is usually expressed with certain patterns in various sentences.This clue provides us a new way for relation extraction that we can learn a similarity metric between two sentences.During training phase,we update the parameters of the model to project similar relation pairs nearer to each other and dissimilar relation pairs farther away from each other.In testing phase,given a new sentence,we calculate the distance between the new sentence and some sentences which relation types are already known by our model.Then,we use k-nearest neighbors algorithm to get the result.In the experiment,we demonstrate the effectiveness of our model on a public data set.(2)To infer latent triplets based on the existing data in knowledge graph,we make full use of the connection of nodes when modeling the triplets.For the first-order connection,we use different weights to distinguish the contribution degree of different relations and combine all the entities in the local connection.Then,we translate it to the tail entity according to the type of relation.(3)Furthermore,for the second-order connection of a node,we merge all the connection information into a three dimension matrix and use fully convolution operations to extract features.If the node have external text information,we use character-level bidirectional LSTM to extract features.Then,we can concatenate these two feature vectors into a single vector.Finally,we infer the tail node for the entity node in a certain relation.Following previous work,we evaluate our model on the tasks of triplet classification and link prediction.The experimental results show that our models are very competitive.Our plans for future work is to investigate new strategies and improve the performance of the knowledge graph completion model.Also,we can make an in-depth study on the domain knowledge graph in the real world.