Research on Named Entity Recognition in the Steel Industry Based on MacBERT

Abstract. In response to the lack of larger-scale and high-quality NER datasets and research on NER in the steel industry, this paper constructs a NER dataset for the steel industry that includes 4835 pieces of data, and annotates four entity categories: device, material, process, and product, A NER research method based on MacBERT_large-BiLSTM-CRF model was built. This method first utilizes the MacBERT model to generate semantically rich dynamic word vectors, then inputs the word vectors into the BiLSTM network model to obtain global features, and finally uses the CRF model to add effective constraints to the test labels to ensure the effectiveness of the generated labels. The model was compared with three other models, and the experimental results showed that the precision of the model was 90.01%, the recall rate was 91.02%, and the F1 value was 90.51%. The recognition performance of the model was superior to the other three models.