Research On Knowledge Representation Learning Based On Entity Description And Entity Similarity

Knowledge Graphs describe the entities and relationships in the real world,which has become the foundation of many intelligent applications.Since the concept of knowledge graph was proposed by Google in 2012,domestic and foreign Internet companies have rushed to develop their own knowledge graph products and applied them to relationship extraction,intelligent search,intelligent question answering,and personalized recommendations,such as Microsoft Bing Satori and Sogou.Cube and Baidu Zhixin etc.The successful construction and application of knowledge graph depends to a large extent on the rationality of the representation method of knowledge graph.A typical knowledge graph is a collection of triples,where h and t represent the head and tail entities,respectively,and the relationship r between h and t.This symbolized representation is not conducive to numerical calculation,and large-scale knowledge graphs usually have high dimensionality,sparseness,and incompleteness.At the same time,due to the explosive growth of knowledge,knowledge graphs need to be continuously updated and completed.The inference algorithm has high calculation complexity and poor scalability,and it is difficult to meet the requirements of large-scale real-time calculation.The traditional representation method takes the one-hot representation as an example.It assumes that all objects are independent of each other,that is,the vector representation of the objects is orthogonal,no matter whether the semantics of the two objects are similar,the semantic similarity calculated by cosine distance is zero,causing loss of semantic information.At the same time,as the size of the object increases,the vector dimension under the one-hot representation also increases,which cannot reasonably represent the triples in the knowledge graph.Therefore,researchers represent triples as dense,low-dimensional vectors,which can effectively alleviate sparsity,and can efficiently calculate the similarity between entities through Euclidean distance and other methods,thereby improving the performance of knowledge acquisition,fusion,and reasoning.Currently,the translation models represented by TransE are typical works,but this type of method models triples from a structural point of view,and generates vector representations by learning triple structure information,overlooking the rich variety of additional information,or failure to mine effective information to enhance the expressive power of the knowledge representation learning model.Moreover,many models treat the knowledge graph as an independent set of triples,ignoring the hidden features of the original graph,and the embedded representation has poor interpretability.For the above-mentioned areas to be improved,this thesis introduces entity description and entity similarity in the process of knowledge representation learning,and proposes the following two methods:(1)Representation Learning with Structural and Descriptive Modeling(DSRL)is proposed.The description information corresponding to the entities in the knowledge graph contains rich semantic information,which can help to improve the data sparsity of the knowledge graph.It is based on the existing entity description text.The entity description information strongly related to the triple entity,and then effectively encode the entity description information,model the knowledge graph representation learning from both the triple structure and entity description,and use experiments to verify the effect of the proposed method.(2)The knowledge representation learning method SimE based on entity similarity is proposed.Inspired by manifold learning,it is believed that similar entities in the original map should also retain similar characteristics in the embedding space.The structural neighborhood of the entity is introduced to measure the similarity of the entities Then,the similarity and Laplace feature map are combined as constraints based on structural representation,which can effectively improve the expressive ability and interpretability of the model,and also verify the effectiveness of the method through experiments.