Abstract: Aiming at extracting effective entity information from unstructured corpus in the field of food safety,a named entity recognition (NER) method based on BERT (Bidirectional Encoder Representations from Transformers) and adversarial training is proposed.NER is a typical sequence labeling problem.At present,deep learning methods have been widely used in this task and have achieved remarkable results.However,there are problems such as difficulty in constructing a large number of sample sets for NER in specific fields like the food field,and inaccurate recognition of proper noun boundaries.To solve these problems,BERT is used to get the word vector,which enriches the semantic representation.To optimize the NER task,adversarial training is introduced,which not only uses the shared information obtained from task training of Chinese word segmentation (CWS) and NER,but also prevents the private information of CWS task from generating noise.The experiment is based on the corpus of two categories,which are Chinese food safety cases and People's Daily news respectively.Among them,the Chinese food safety cases data set is used to train the NER task,and the “People's Daily” news data set is used to train the CWS task.We use adversarial trainingto improve the precision of the NER task for entity recognition (including name,location,organization,food name and additive).Experimental results show that the proposed method's Precision rate,Recall rate and F1 score are 95.46%,89.50% and 92.38% respectively.Therefore,this method has a high F1 score for Chinese NER task,where the boundary of a specific domain is indistinct.

Key words: Adversarial training, BERT, BiLSTM, Food field, Named entity recognition