BERT-Based Chinese NER with Lexicon and Position Enhanced Information Adapter

Recently, many approaches have performed Chinese NER using pre-trained models adding lexicon information, which has been proven to be more effective. A recent study LEBERT improves performance by fusing lexicon feature between the beginning layers of BERT. However, LEBERT ignores the po-sition information of words in sentences. In this paper, we propose Lexicon and Position Enhanced BERT (POSLEBERT) for Chinese NER, which fuses lexicon feature and position feature of words into BERT layers to solve the problem that LEBERT doesn't use position feature of words. Compared with BERT-based methods, POSLEBERT can fully fuse character feature, lexicon feature and position feature of words. Experiments on four datasets show POSLEBERT out-performs other BERT-based or Lexicon-based models on four datasets and achieve state-of-the-art results.