Research On Recognition Method Of Aerospace Intelligence Named Entity Based On Pre-training Model

Aerospace intelligence entity recognition is an subtask of aerospace intelligence information extraction.The biggest problem facing the aerospace intelligence information extraction task is the lack of annotated corpus.Therefore,this paper takes the problem of the lack of annotated corpus in the aerospace intelligence field as an entry point.On the one hand,it is proposing a fusion of multi-source heterogeneous knowledge to build a knowledge base,knowledge base-driven annotation data set.On the other hand,the pre-training model is used for transfer learning to solve the problem of less labeled data.This paper mainly studies Aerospace Intelligence Entity Recognition(AIER).The main research contents of this paper are as follows:1.Propose corresponding extraction and fusion methods for the aerospace knowledge contained in databases,books,and the Internet,and build aerospace intelligence knowledge base.Use the knowledge base-driven heuristic labeling algorithm to label the data set.2.Construct an aerospace intelligence entity recognition model based on statistical machine learning methods.In this paper,HMM and CRF models are used for aerospace intelligence entity recognition tasks,and it is found that the CRF model has better effects in label prediction.3.Combine the representation learning ability of deep learning and the structured prediction ability of the CRF model and apply it to the mission of aerospace intelligence entity recognition.A Bi LSTM-CRF model is constructed,which can achieve 91.88%accuracy,93.64% recall,and 92.75% F1 value.4.Start with the problem of small aerospace intelligence entity recognition data sets,use the pre-training model trained on a large-scale Chinese corpus,and apply it to the aerospace intelligence entity recognition task through the method of transfer learning.Constructed the BERT-CRF model and the ALBERT-CRF model respectively.Among them,the BERT-CRF model can achieve an accuracy of 93.68%,a recall rate of 97.56%,and an F1 value of 95.58%.Compared with the Bi LSTM-CRF model,the precision has increased by 1.8%,the recall has increased by 3.92%,and the F1 value has increased by 2.83%.5.This paper builds a Chinese aerospace intelligence entity recognition software system based on a pre-training model,which can quickly and accurately identify aerospace entities from aerospace intelligence.