Research On Knowledge Representation Learning Based On Quantum Theory And Hyperbolic Space Theory

As a very important tool in the information age,knowledge graph has been playing an increasingly important role in information retrieval,knowledge questions and answers,recommendation system and many other aspects.With the advent of the information age of big data,it is difficult to meet people's demand for accurate information with increasing data,so the construction of knowledge graph provides convenience for people.However,there are some difficulties in the storage and representation of large-scale knowledge graphs.In recent years,various presentation learning techniques have attracted researchers' attention and contributed to the representation of large-scale knowledge graphs,but there are still many technical problems.Knowledge graphs are generally represented in the form of triples(entities,relations,entities).The existing methods generally represent these entities and relations in the form of vectors,and make full use of the characteristics of word vectors to effectively represent all kinds of knowledge triples in the knowledge graphs.Distributed presentation is a relatively efficient presentation learning technique in recent years.The classical translation model TransE method can effectively solve problems such as data sparsity,but the representation of complex relationships is not ideal,so it is particularly important to study a more scientific and efficient representation learning method.Based on the existing research work,this paper further improves the research work from the following two aspects:(1)Considering the time-sensitive nature of knowledge in the knowledge graph,the triple relationship may change with time.In recent years,quantum theory has been significantly applied in machine learning and other fields.A knowledge representation method based on quantum theory named Q-TransX is proposed,combined with quantum polymorphism to solve dynamic problems in the knowledge graph,uses the word embedding representation learning method to train quantum embedding,and the experimental results verify the effectiveness of the method.(2)Hyper-TransE,a hyperbolic embedding method based on hyperbolic space theory,is proposed.This method can use hyperbolic space to capture the characteristics of knowledge information with hierarchical structure,and there are a lot of hierarchical relationships in the knowledge or relationship in the knowledge graph.Hyperbolic embedding combined with European-style embedding not only captures this hierarchical relationship,but also greatly reduces the embedding dimension,solves the dimensional disaster problem faced by word vectors,and greatly improves the quality of representation learning.Experiments have proved The effectiveness of this method.Based on quantum theory and hyperbolic space theory,Q-TransX method and Hyper-TransE method are proposed to solve some problems in knowledge representation learning.Experiments on some classical data sets show that the performance of our method has been improved to some extent.