Knowledge Graph Link Prediction Based On Multi-Task Learning

A knowledge graph is a structured method of describing the real world.However,due to the complexity of the real world and the diversity of knowledge,it is inevitable that some information will be omitted in the construction of a knowledge graph.An effective method for solving this problem is link prediction,which can use machine learning algorithms and natural language processing techniques to infer missing information by leveraging existing knowledge graph information and semantic relatedness.Although the existing knowledge graph link prediction research has achieved certain results,there are still the following problems:(1)Most of the traditional link prediction models are based on the modeling of entities in the knowledge graph,ignoring the semantic association between different relationships.Therefore,the relationship information is not fully updated,resulting in inaccurate prediction results.(2)Some studies have overemphasized the interaction between entities and relationships,ignoring the original translation characteristics in triples,resulting in incomplete triple feature mining and affecting the accuracy and effectiveness of prediction.In this dissertation,the specific research on the above problems is as follows:Firstly,in previous studies,only entities in the knowledge graph are usually used as modeling objects,ignoring the semantic associations between different relationships.In order to solve this problem,this dissertation uses a graph convolutional neural network to capture the dependencies and relationship topology information between nodes,and uses a convolutional network to decode the triples to complete the link prediction task.Specifically,this dissertation proposes a new graph convolutional neural network method based on relationship topology information.This method defines the relationship between adjacent relationships as three topological structures,and models these structures when constructing graph convolutional neural networks.For different topological structures,an independent update strategy is used to update the information of the relationship.Secondly,the link prediction model based on convolutional neural network destroys the translation characteristics in triples.This dissertation proposes a new representation method.Based on the graph convolutional neural network as the encoder,the interaction and translation characteristics between entities and relationships in the captured triples are considered to improve the performance of the model.In this dissertation,the gated network is used to control the influence of the shared layer on different tasks,and the translation characteristics are effectively integrated into the convolutional neural network by expanding the parameter space of the model,thereby improving the performance of the model.In this dissertation,experiments are carried out on public datasets such as UMLS,WN18RR and FB15k-237 for link prediction tasks.Among them,this dissertation conducts an experiment of graph convolutional neural network based on relational topology information on FB15k-237 dataset,and conducts an experiment of multi-task link prediction algorithm based on graph convolutional neural network on UMLS,WN18RR and FB15k-237 datasets.Compared with the baseline model,the model proposed in this dissertation has better experimental results on MRR,MR and other indicators,which proves the significance and value of the model proposed in this dissertation.The experimental results of ablation experiments prove the effectiveness of the proposed modules and the rationality of the model structure.