A weakly supervised back-labeling-based method for named entity recognition in automotive manufacturing: MBIC-NER

Aiming at the problems of lack of corpus, lack of entity annotation dataset and fragmented data in named entity recognition in the automobile manufacturing field, a new named entity recognition method is proposed: MBIC-NER. The method firstly constructs an automobile manufacturing annotation dataset by using the idea of weakly supervised back annotation, and generates word vector sequences of the input sentences by using MacBERT pre-training model in order to better understand the context and semantic associations, and improves the accuracy and effect of text processing in the field of automobile manufacturing; Secondly, BiLSTM model and IDCNN model are constructed to perform feature extraction on the input word vector sequences to form a multilevel feature representation; Finally, the character sequences are continued to be decoded through the CRF, and the corresponding label of each character is obtained. The result show that the F1-score of MBIC-NER for named entity recognition in the self-constructed automobile manufacturing domain dataset reaches 87.03%, which is better than the other named entity recognition methods in comparison.