Research On Entity Action And Relation Extraction Based On Knowledge Graph Construction

The construction of knowledge graph is to store all kinds of information such as related entities and related information in the required fields in the form of RDF tuples in a graph base with efficiency and storage,and achieve a specific path according to requirements.For the construction of knowledge graph,one important thing is information extraction in the corresponding field.Information extraction is mainly about the extraction of entity attributes and relationships,usually a sentence level extraction process.Most of the sentences are unable to extract attributes and relationships.In KBP competition,it considers attributes as alias name,age,birth place and so on.In order to improve the utilization of text,we consider two aspects of optimization;The sentence which cannot be filled into slots but involved entity,should be reverted to information to save;According to new requirements,a relation extractor can be constructed quickly.Therefore,the main research direction of this dissertation is to improve the utilization of unstructured text,enrich the knowledge graph,and provide the data base for it.The main work includes: An action extraction framework based on open domain extraction is designed.The sentences involving entities but cannot fill into slots,which are used to indicate actions initiated or accepted by a entity,or to indicate interactions between multiple entities.In this dissertation,such information is regarded as human actions.The extraction and storage of human action can enrich the knowledge graph,provide the corresponding database for knowledge graph,and enrich the associated relationship between entities.a specific relation extraction framework based on user assisted correction is designed.When a new relationship is required to extract from sentences,the annotation of the corresponding data can be guided through the framework,and a new classifier can be constructed iteratively based on a small number of tagging data and user's auxiliary process,so as to improve the extraction performance of the specific relationship.a Bi-LSTM network relation extraction model based on physical and semantic feature is implemented.Adding physical and semantic features to the input of LSTM network can change the classification process from sentence level to entity pair level,and get different classification results for different entities in the same sentence,which effectively improves the accuracy rate.The framework and model described in this paper prove the feasibility of enriching the knowledge graph through the aspects mentioned above in terms of theory and experiment.