Food Named Entity Recognition with BERT and Adversarial Training

To efficiently identify food entities within unstructured text, this paper introduces a named entity recognition (NER) approach leveraging BERT (Bidirectional Encoder Representations from Transformers) combined with adversarial training. As a common sequence labeling task, NER has made significant progress by applying deep learning techniques. But in some specific fields, such as the food field, NER still faces challenges such as constructing large sample sets and accurately identifying the boundaries of specialized nouns. We use BERT to obtain word vectors, enriching the semantic representation. We introduce privacy information noise in the Chinese word segmentation (CWS) task and use adversarial training to improve the accuracy of entity boundary recognition by using the shared information between CWS and NER. We conducted experiments using Chinese food safety cases and the People's Daily corpus. The training dataset for the NER task used Chinese food safety cases, and the training dataset for the CWS task used the People's Daily corpus. To improve the recognition accuracy of different entity types in the food field, including food names, additive names and related place names, organization names, our proposed method employed adversarial training. The experimental results indicate that the method attains an accuracy of 95.78%, a recall of 89.96%, and an F1 score of 93.47%. Therefore, this method proves effective for NER in the Chinese food domain, particularly in identifying entities with ambiguous boundaries.