Research On Military Named Entity Recognition Method Based On Pre-training Language Model

With the advent of the information age,the problem of "overload" of military textual information has increased,which has placed a heavy burden on intelligence analysts.Using the information extraction technology in natural language processing to intelligently extract and present the valuable information is a solution that has attracted much attention.The importance of named entity recognition technology as a cornerstone in the field of information extraction is self-evident.Based on the task of military named entity recognition,this paper conducts research on deep learning methods based on pretrained language models.Because of the lack of relevant corpora in the military field,for research needs,this article first constructed a military named entity recognition dataset using military news text as the data source.The dataset has a scale of 900,000 words,covering six named entities of person,organization,location,time,rank and position,and weaponry.In the follow-up,this article carries out experimental research and comparative analysis of several methods around this dataset.In this paper,BiLSTM-CRF is used as the benchmark model,and the BERT-CRF model and the BERT-BiLSTM-CRF model are constructed respectively,and their performance is compared.Experiments show that the performance of the BERTBiLSTM-CRF model is the best,and the BERT model and the BiLSTM model can complement each other.The overall F1 value of the BERT-BiLSTM-CRF model on various entities is 90.66%,which is 6% higher than the BiLSTM-CRF model.The BERT model is based on word vectors and does not consider word information.How to properly integrate word information into the word vector-based model to achieve a gain effect is also a promising research direction.This paper verifies the effectiveness of this idea in the field of military named entity recognition by combining the BERT model with the FLAT(Flat-Lattice Transformer)model.Then,through comparative experiments on the BERT-base,BERT-wwm and Ro BERTa models,it is found that the model based on the combination of BERT-wwm and FLAT performs best overall,with an overall F1 value of 91.14% on various entities.