Research And Implementation Of Knowledge Graph Construction Method Oriented To Military Text

There are massive amounts of data and valuable information and knowledge on the Internet.How to organize existing data and information has become a problem that needs to be solved urgently.Knowledge graphs,as a way of storing knowledge today,can be combined with specific fields to provide refinement knowledge for this field.This paper aims at a large amount of military text data on the Internet,using information extraction technology to extract valuable information,and constitute a military knowledge graph,laying the foundation for information construction and knowledge reasoning in the military field,this paper starts research from three aspects,The domain word discovery,named entity recognition and relationship classification technology.The specific work is as follows:First,this paper studies and improves the domain word discovery method of combining statistics with clustering,using the method of combining point mutual information and branch entropy statistics to evaluate the internal solidification degree and external degree of freedom of the constituent words,and then clustering through K-means method separates the common words and domain words in the candidate words to form the domain word lexicon in the military field.Secondly,this paper studies and improves the named entity recognition method based on Lattice LSTM.First,a rule-based method is used to pre-label unlabeled military texts,and a named entity recognition dataset in the military field is constructed.Then we studied a named entity recognition model based on deep learning,which uses a presentation layer based on the BERT model,a context coding layer based on the Lattice LSTM model and a label decoding layer based on conditional random fields.The experimental results show that this model performs better than other models on the named entity recognition data set of People’s Daily.The model is applied to the military named entity recognition data set,which realizes the named entity recognition method in the military field,saving 95% of manpower and material resources.Thirdly,this paper proposes a few-shot relationship classification method based on co-attention mechanism and dynamic routing(CADR),mainly to solve the problem of lack of a large number of relationship classification datasets in the military field.The method is divided into three parts,The encoding layer of the attention mechanism,the instance aggregation layer and the relationship matching layer,the experiment shows that the method performs better on the 5-way-10-shot problem on the Few Rel dataset than other few-shot learning models,and has been verified by experiment The co-attention mechanism enhances the information of the samples in the training set,and verifies the improvement effect of the distance measurement loss function in the loss function on the model.Finally,based on the above research,this paper designs and implements a knowledge graph construction system in the military field.The system is divided into three modules: data acquisition and storage module,data analysis module and result visualization module.The data acquisition and storage module is used to obtain military texts on the Internet and store the intermediate data generated by the data analysis module.The data analysis module includes the training and use of domain word recognition,named entity recognition models,and relationship classification models.The result visualization module incorporates system functions Provided to users as a service,and can query and analyze the data in the knowledge graph,and the results are displayed in the form of a knowledge graph.