Research On Text Based Entity Typing In Knowledge Graph

With the rise of the Internet and the arrival of the era of big data,the scale of data on the Internet has grown exponentially,and these data contain massive knowledge.In order to make use of this knowledge effectively,knowledge graph comes into being.As one of the most important knowledge representation forms,knowledge graph is a kind of large-scale semantic network in essence,including entities,concepts and various semantic relationships among them.The entity typing of knowledge graph is an important task in the construction of knowledge graph.Its goal is to establish the genus relationship between entity and concept,and the determination of the relationship is helpful for downstream applications such as intelligent search,intelligent question-answering,intelligent recommendation,etc.Entities in knowledge graph include both structured description and unstructured text description.The current entity typing methods of knowledge graph are mainly based on the structured description feature of entities.However,due to the incompleteness of knowledge graph,the entities inevitably lose some structured features,resulting in the entity cannot get complete typing results,which reduces the entity typing effect.As an important supplement to the current entity typing methods,this paper studies the entity typing of knowledge graph based on the text description.However,the current text-based entity typing methods in knowledge graph still have many shortcomings,including:(1)the current methods do not make full use of the entity’s text description information;(2)The current approaches fail to make full use of the semantic information of concepts.In this paper,the shortcomings of existing methods are solved,and proposes corresponding solutions.The main contents and contributions of this paper are summarized as follows:(1)To solve the problem that the current methods fail to make full use of the text description information of entities,this paper constructs a heterogeneous graph based on all sentences in the text,in which words and sentences are regarded as different types of nodes,and the semantic relations between them are regarded as different types of edges.On this basis,an entity typing method based on entity aware heterogeneous graph neural network is proposed.The model consists of three parts,namely entity-aware encoder,context encoder and concept decoder.Firstly,this paper uses an entity-aware encoder to encode sentences and entity names to initialize the representation of each word and sentence in the heterogeneous graph.Secondly,a context encoder based on the heterogeneous graph attention network is used to obtain the context node representation of each word and sentence.Finally,a concept decoder based on a multi-layer perceptron network is used to obtain a set of concepts for each entity.(2)To solve the problem of the current methods fail to make full use of the semantic information of concepts,this paper constructs a heterogeneous graph based on the description text and concept,in which the text and concept are regarded as different types of nodes.On this basis,an entity typing method based on concept-enhanced heterogeneous graph neural network is proposed,and the semantic representation of concepts and their co-occurrence relationship and hierarchy are considered.The model consists of four parts,namely entity encoder,concept encoder,concept enhancement encoder and concept decoder.Firstly,to initialize the representation of each entity and concept in the heterogeneous graph,this paper introduces entity encoder and concept encoder respectively to encode the text of the entity and concept.Secondly,a concept enhancement encoder based on a heterogeneous graph attention network is used to obtain an enhanced representation of each entity and concept.Finally,a concept decoder based on a multi-layer perceptron network is used to obtain a set of concepts for each entity.(3)Due to there is no publicly available text-based knowledge graph entity typing data set at present.This paper constructs a dataset based on DBpedia,an English knowledge graph.On this basis,two methods are proposed to evaluate this paper.The overall experimental results demonstrate the effectiveness of the two proposed methods.At the same time,the effectiveness of each module of each method is proved by the ablation experiment and case analysis.