Research On Information Extraction Methods For Information Equipment Knowledge Graph Construction

With the development of big data technology,various fields are beginning to focus on the role of data drivers,including the military.The importance of information is now increasingly valued in warfare.Information equipment updates and developments have derived a large amount of text data.In information warfare,automatic access to key information of information equipment can assist in the process of analysis and decision making;however,useful information is submerged in a large amount of unstructured text,and it is difficult for limited manual effort to efficiently process the large amount of text data.Information extraction technology is the basis for building the knowledge graph of information equipment,which can realize efficient useful information extraction,including named entity recognition and relation extraction.The construction of the knowledge graph of information equipment can analyze these data with complex relationships more efficiently.Therefore,this dissertation studies the construction of more effective information extraction model and uses the information extraction model for the construction of information equipment knowledge graph.The main research work of this dissertation is mainly in the following aspects.First,for named entity recognition,this dissertation proposes a head-to-tail linker to locate and extract entities.When extracting entities in text,there are often nested structures between different entities.The head-to-tail linker is a boundary-based method that can enhance the association between entity boundaries and entity types while dealing with the nested entities.The head-to-tail linker extracts all possible entity heads as the starting point and explores the corresponding entity tails in the space of different entity types.To learn the span information of entities and avoid losing the entity length limit,the relative position encoding embedded in the linker.In this dissertation,experimental results are compared on three publicly available datasets and different benchmark models to demonstrate that the head-to-tail linker is able to extract nested entities better.Second,for relation extraction,this dissertation proposes a relation-aware mechanism to assist different models to extract relational triples.At present,a popular problem in the field of relation extraction is the overlapping problem of relational triples.Some current models,such as Cas Rel,deal with the problem of overlapped triples well,however,these models focus on solving the overlapping of entities between different triples,and correctly classifying relations is also a key issue.In this dissertation,we encode the relation label information as vectors and analyze the association between text and relation labels by means of attention,i.e.,we reinforce the mapping of input text to the relationship category space in the task space by means of relation-aware mechanism.To effectively demonstrate the effectiveness of the approach,this dissertation embeds the relation-aware mechanism into three relation extraction models SMHMA,ETL-Span and Cas Rel,and conducts extensive comparison experiments on two publicly available datasets.The experimental results demonstrate that the relation extraction models embedded with the relationship-aware mechanism can better locate the relational triples compared with the original models.Third,this dissertation constructs a information equipment knowledge graph in the electromagnetic domain to assist the military analysis and decision-making process.The construction of the information equipment knowledge graph mainly consists of two key parts: training information extraction models and importing the data extracted by the models into the graph database.For training information extraction models,the main work of this dissertation is to construct information equipment named entity recognition dataset and information equipment relation extraction dataset to train the corresponding information extraction models,which can obtain information equipment entities and triples from unstructured and semi-structured texts.For building the graph database,this dissertation imports the extracted data into the graph database through Neo4 j and provides two ways to query the information equipment knowledge graph.