Research On Representation And Reasoning Methods Of Structured Knowledge

Structured knowledge is the information extracted from massive data,which is highly generalized and organized,and stored and utilized in a unified structural format.The rep-resentation and reasoning of knowledge is to represent the knowledge graph—the typical structured knowledge as a continuous numerical vector directly processed by the com-puter,and apply it to downstream tasks to mine more implicit information.Currently,knowledge representation and reasoning have played a significant role in various fields,driving the development direction of artificial intelligence from being solely data-driven to a combination of“data + knowledge”driving force.However,existing knowledge representation and reasoning work is not yet perfect and still has some shortcomings.First of all,the lack of structure information in the pro-cess of knowledge representation makes it difficult for the learned knowledge embedding to reflect the graph structure information of the knowledge graph.Secondly,the semantic association in the knowledge graph is implicit,making it difficult for the learned knowl-edge embedding representation to reflect its potential semantic association information?Then,the filtering of knowledge noise in the knowledge reasoning process is complicated,and it is difficult to accurately filter the knowledge noise that is irrelevant to downstream tasks? Finally,it is difficult to integrate knowledge features in the process of knowledge reasoning,and it is difficult to efficiently integrate downstream tasks with knowledge graphs.Therefore,this dissertation conducts a systematic study focusing on the short-comings of existing works.The main research contents are as follows:(1)To address the issue of missing structure information during the knowledge repre-sentation process,a heterogeneous graph embedding model based on structure information aggregation has been proposed.This model realizes accurate learning of heterogeneous graph embedding representation by effectively aggregating local structure information and global structure information of heterogeneous graphs.The model first decomposes the het-erogeneous graph into different types of meta-paths and several corresponding meta-path instances,and uses the attention mechanism to capture the local structure information of the heterogeneous graph.Furthermore,an homogeneous graph conversion mechanism has been designed that ignores the semantic information of nodes and relations in the heteroge-neous graph and converts it into a corresponding homogeneous graph,thereby extracting global structure information.Finally,the captured local and global information are effec-tively aggregated to obtain the final embedding representation of the heterogeneous graph.Experimental verification based on various types of heterogeneous graph datasets shows that the proposed model outperforms existing heterogeneous graph embedding models in terms of node classification,node clustering,and link prediction tasks.(2)To solve the issue of implicit semantic associations within the knowledge graph,an entity-relation interaction graph neural network model has been proposed.The model effectively combines traditional knowledge embedding algorithms with graph neural net-work techniques,explicitly incorporating relation embeddings.Firstly,the model aggre-gates the information of multi-hop neighbor nodes into corresponding entity embeddings by constructing auxiliary edges.Then,by constructing a matrix of semantic similarity between relations,the model aggregates the information of semantically similar relations and their associated entity information into corresponding relation embeddings.Finally,the learned entity and relation embeddings are evaluated using the scoring function of traditional knowledge embedding algorithms.Experimental results on multiple knowl-edge graph datasets show that the proposed model outperforms most existing knowledge embedding models in link prediction and node classification tasks.(3)To address the complex issue of filtering knowledge noise during the knowledge reasoning process,a knowledge recommendation model based on contrastive learning mechanism has been proposed.The model is designed for the typical reasoning scenario of recommendation systems.It utilizes a contrastive learning mechanism to accurately filter out task-irrelevant noise information introduced from the knowledge graph,thereby improving the accuracy and efficiency of reasoning.Firstly,the model performs data aug-mentation on the knowledge graph and historical interaction data to filter out irrelevant knowledge and introduces item-item relations,constructing knowledge view and interac-tion view separately.Then,an relation-aware graph neural network effectively encodes the knowledge view.Finally,by designing a contrastive learning-based loss function,cross-view associations are established to realize a knowledge-based recommendation model.The proposed model has been experimentally validated using various recommen-dation datasets from different scenarios.It outperforms existing recommendation models in terms of Top-20 recommendations.(4)To address the challenge of integrating knowledge features in knowledge rea-soning,two typical visual reasoning tasks,fine-grained image classification and visual relationship detection,have been addressed with two proposed reasoning models: a rea-soning model based on knowledge graph embedding fusion and a reasoning model based on knowledge feature correlation mapping.Since the raw data of external tasks already contains abundant information,it is essential to effectively integrate the key features in-herent in the data with the introduced knowledge features in order to obtain higher-order information.However,existing methods only rely on simple operations such as feature concatenation and addition to fuse knowledge features.Therefore,the proposed model in this dissertation leverages the image label information from visual tasks and publicly avail-able common sense knowledge graphs to construct a visual reasoning knowledge graph.Then,two fusion methods are adopted: knowledge feature interaction fusion and knowl-edge feature correlation mapping,to efficiently integrate knowledge features with visual features,thereby enhancing the performance of visual reasoning tasks.Experimental re-sults on visual reasoning datasets show that both proposed models outperform existing visual models in terms of reasoning accuracy.