Research on effects of pre-trained language models in medical named entity recognition

Named entity recognition (NER) is a step stone for numerous downstream applications, and medical NER is an important part of NER. Prior studies have applied various pre-trained language models (PLMs) to medical NER, but they ignore to systematically investigate cons and pros of these PLMs. In this paper, we investigate cons and pros of prevalent PLMs in medical NER. To be specific, we first pre-train three PLMs (i.e., word2vec, GloVe and ELMo) from scratch and fine-tune Chinese BERT model with 300k entries of real-world Chinese Electronic Medical Records. Then, we combine above PLMs with BiLSTM-CRF to evaluate effects of these PLMs. Experimental results on CCKS2019 dataset show that context-dependent PLMs (ELMo and BERT) significantly outperform con-text-independent PLMs (word2vec and GloVe) in medical NER, by up to 4.98% absolute F1 gains. Moreover, our best model achieves new state-of-the-art results on this benchmark dataset. Furthermore, we make additional analyses from perspectives of time and space complexity.