Construction Of Knowledge Graph In Robotics Based On Multi-source Text Analysis

Artificial intelligence and robotics are important fields of current technological development.Patents and other scientific and technical literature reflect the progress of fundamental research and technological innovation.Applying natural language processing technology,and combining the robots and other important technical development fields with patents and other scientific and technical documents for in-depth research,can help the practitioners in these domains to efficiently find solutions to product design problems,and to realize domain knowledge mining,subject discovery and relevance evaluation,and evolutionary trend analysis.This is of great significance for understanding the interaction and penetration of different scientific and technological fields and discovering potential business opportunities.Knowledge graph is a kind of knowledge representation model proposed by Google for optimizing search engines,which reveals the relationship between entities.Compared with traditional knowledge base,knowledge graph can help people to obtain the logical relationship between required knowledge more quickly and effectively and to realize the intelligent reasoning among knowledge.Based on the analysis of existing knowledge graph construction methods,this thesis studies the methods of constructing vertical domain knowledge graphs based on multiple data sources,and proposes and the design schemes for developing a knowledge graph application system.The main contributions of this thesis are as follows:(1)Aiming at the lack of field training corpus,a training corpus construction method based on remote supervision is developed.First,a tree structure diagram of domain knowledge is constructed.Then the structured information in the robotics domain is collected from Baidu Encyclopedia by the web crawler technology and taken as the initial three-tuple.In addition,the remote supervision technique is used to automatically obtain the training corpus.(2)Two extraction methods are used for different types of entity relationship extraction tasks.For entity recognition tasks,a heuristic-based rule screening method and a K-Nearest Neighbor(KNN)classification screening method are used.For entity relationship extraction,a syntactic rules-based entity relationship extraction method is used,together with a relationship extraction model based on distant supervision.(3)Using the robot knowledge three-tuple obtained by the above methods as a data source,a knowledge storage scheme based on the Neo4 j graph database is proposed to store the threetuple.Furthermore,query and display of the robot domain knowledge graph are implemented via the visualization platform of the graph database.(4)Based on the constructed knowledge graph,the upper-level domain application is designed,entity recognition for paragraph text is conducted,and the key information of the text will be quickly mastered.A tree-structured diagram is constructed for the entities in the domain and the key information,e.g.,the problems and solutions,is displayed.An entity relationship query module is set up to quickly return the query content,and a question and answer module to is set up to return the answers for the questions as well as related patent recommendations.