Research On Knowledge Representation Learning Algorithms For Knowledge Graph Completion

Knowledge graph contains structured knowledge and has become the cornerstone of many artificial intelligence applications.In order to expand the scale of knowledge graph,knowledge graph completion task has gradually received researcher's attention.The knowledge representation learning algorithm is one of the best solutions for knowledge graph completion task.This kind of algorithm not only has strong generalization ability,but also has good scalability.However,after in-depth analysis and research,we find that existing knowledge representation learning algorithms still have many shortcomings,and improve them from three perspectives:1)The dynamic knowledge graph completion task for new entities can simultaneously expands knowledge graph's entity set and triple set,which has good practical significance.However,existing knowledge representation learning algorithms cannot handle this task well.Therefore,this paper proposes a novel knowledge representation learning algorithm according to the essence of this task.In addition to utilizing neural networks to model description information,we also design two encoders for constructing structural vectors.Among them,the final vector of the entity is derived from the combination of the textual vector and the structural vector.2)Most knowledge representation learning algorithms require random generation of negative samples during the training process,and the quality of negative samples will seriously affect the generalization ability of the algorithm.Therefore,this paper analyzes the stability and advantages of existing negative sample generation strategies through sufficient experiments.In addition,we design two new negative sample generation strategies.3)Knowledge representation learning algorithm can obtain entity's vector through end-to-end training.Actually,these vectors can be regarded as high-order nonlinear features.Based on this idea,this paper proposes a novel data prediction framework,which is based on knowledge representation learning algorithm.Among them,because the goal of data prediction tasks is usually specific,there is a gap with the globally oriented knowledge representation learning algorithm.Therefore,we propose two local knowledge representation learning algorithms based on target relationships.We conducted experiments on the above three aspects and compared our algorithms with the corresponding baseline.The experimental results show that our algorithm achieves good results in the static and dynamic knowledge graph completion,which demonstrates that the semantics of entities and relationships are better characterized.In addition,the negative sample generation strategy proposed in this paper is superior to the existing strategy at HITS@10.At the same time,we also found that the stability of different strategy's results is mainly affected by the dataset.And the data prediction framework proposed in this paper is better than most of baseline algorithms,which shows that it is a feasible solution to do data prediction tasks from the perspective of knowledge representation learning algorithms.