| At present,the novel coronavirus pneumonia epidemic in China is sporadic.In order to quickly achieve the key data in the survey work,and help early detection of existing cases,close contacts and close contacts,and cut off the transmission chain,this paper studies the automatic identification of key information by using Named Etity Recognition（NER）technology in the field of artificial intelligence.Provide technical support for the rapid work of flow control personnel in the later stage.Therefore,this novel coronavirus pneumonia novel coronavirus pneumonia track information is the main data source.Starting from the corpus of Track information Entity Recognition of COVID-19 confirmed cases（TiERoCOVID-19）,we use artificial annotation to label small datasets,and on the other hand,we solve the problem of small datasets by pre training the transfer learning of language models.Through the above two main aspects,the named entity recognition technology is applied to the TiERoCOVID-19 task.This paper completes the whole process from labeling corpus to entity recognition and application system construction.The main research contents are as follows:1.The published track information of COVID-19 confirmed cases was collected and the data was tagged use the experience of other fields for reference by manual marking.A named entity recognition data set based on the track information of COVID-19 confirmed cases was constructed.2.An entity recognition model based on statistical machine learning method is constructed.This paper constructs HMM（hidden Markov models,HMM）and CFR（conditional random fields,CRF）models based on the classical statistical machine learning model,and applies them to the TiERoCOVID-19 task.The experiment shows that CRF model has better effect in the structural prediction of labels.3.Combining the structural prediction ability of CRF model with the expressive learning ability of deep learning,the BiLSTM-CRF model is constructed.The BiLSTM-CRF model can achieve 91.39% accuracy,90.03% recall and 90.87% F1 value on TiERoCOVID-19 data set.4.In order to solve the problem of small data set of TiERoCOVID-19 task,by using the transfer learning ability of pre training language model,an entity recognition model based on pre training language model BERT（Bidirectional Encoder Representations from Transformers,BERT）is constructed and applied to TiERoCOVID-19 task.This paper combines the BERT model with the BiLSTM-CRF model.The accuracy,recall and F1 values of the BERT-BiLSTM-CRF model on the TiERoCOVID-19 data set are 93.29%,93.18% and 93.22% respectively,which is greatly improved compared with the results of the BiLSTM-CRF model.At the same time,inspired by the attention mechanism,this paper continues to introduce the multi head attention mechanism based on the BERT-Bi SLTM-CRF model,and the effect of entity recognition has been further improved.The accuracy,recall and F1 value reach 96.88%,97.55% and 97.21%respectively.In addition,try to change the number of attention heads in the introduced multi head attention mechanism.It is found that the best experimental results can be achieved when the number of attention heads is set to 8.5.A physical recognition system based on pre training language model is built.The system can automatically identify key entities from the unstructured track information entity recognition of COVID-19 confirmed cases. |
