On Cross-Domain Pre-Trained Language Models for Clinical Text Mining: How Do They Perform on Data-Constrained Fine-Tuning?

Pre-trained language models (PLMs) have been deployed in many natural language processing (NLP) tasks and in various domains. Language model pre-training from general or mixed domain rich data plus fine-tuning using small amounts of available data in a low resource domain demonstrated beneficial results by researchers. In this work, we question this statement and verify if BERT-based PLMs from the biomedical domain can perform well in clinical text mining tasks via fine-tuning. We test the state-of-the-art models, i.e. Bioformer which is pre-trained on a large amount of biomedical data from PubMed corpus. We use a historical n2c2 clinical NLP challenge dataset for fine-tuning its task-adapted version (BioformerApt), and show that their performances are actually very low. We also present our own end-to-end model, TransformerCRF, which is developed using Transformer and conditional random fields (CRFs) as encoder and decoder. We further create a new variation model by adding a CRF layer on top of PLM Bioformer (BioformerCRF). We investigate the performances of TransformerCRF on clinical text mining tasks by training from scratch using a limited amount of data, as well as the model BioformerCRF. Experimental evaluation shows that, in a \textit{constrained setting}, all tested models are \textit{far from ideal} regarding extreme low-frequency special token recognition, even though they can achieve relatively higher accuracy on overall text tagging. Our models including source codes will be hosted at \url{https://github.com/poethan/TransformerCRF}.