CLASSIFICATION ATTENTION FOR CHINESE NAMED ENTITY RECOGNITION

The Chinese character-based model, such as BERT, has achieved remarkable success in Chinese named entity recognition (NER). However, such model would likely miss the overall information of the entity words. In this paper, we propose to combine priori entity information with BERT. Instead of relying on additional lexicons or pre-trained word embeddings, our model has generated entity classification embeddings directly on the pre-trained BERT, having the merit of increasing model practicability and avoiding out of vocabulary (OOV) problem. Experiments show that our model has achieved state-of-the-art results on 3 Chinese NER datasets.