Research On Entity Group Profiling In Knowledge Graphs

Knowledge Graphs are graph-structured knowledge bases storing the semantic network of the factual information about real-world entities.At present,the number and scale of Knowledge Graphs are showing explosive growth,each entity description in knowledge graph was associated with more and more facts,which is unacceptably when users want to quickly compare and understand differences between entities.Group profiling is a high practical method to solve this problem,which attempts to replace long description information with concise labels and help users understand and compare differences between groups intuitively and quickly.Traditional group profiling methods require the related field experts to fully understand the data and set the label library.However,the performance of this method is limited when faced with a brand new and huge amount of data.1)The manpower cost.This knowledge engineering usually needs to pay a high human cost,making artificial labeling for large-scale knowledge graphs infeasible;2)Knowledge dynamics.When the scale or domain of knowledge is updated,it is difficult for experts to capture the characteristics of the new knowledge timely and efficiently,resulting in the inability to update the label library quickly and accurately.To solve these problems,proposed an automatic label generation model.Based on the automatically generated label library,further proposed a group profiling model based on attention-based graph neural networks.The proposed method automates the whole process of group profiling.The main research contents of the thesis are as follows:(1)The automatic label generation method is proposed.Analyze the graph-structured knowledge bases and set the types of entity labels.Based on the mathematical statistical inference,discretization of continuous attributes,heuristic rule filtering are used to construct label library automatically and solve the problem of human-defined label library.(2)The group entity profiling based on graph attention neural network is proposed.Applied entity embedding and attention-based graph neural networks.Based on the label library,the topk representative and distinguishing labels are selected,and depict the importance of each label accurately in the groups.(3)Constructed the entity group profiling system of knowledge graphs.Realized the functions of dataset viewing,construct label library automatically,group profiling and so on.The effectiveness of the proposed methods is verified experimentally on real datasets.Experimental results show that the method this paper proposed has improved Hit Rati and average precision.