Inexpensive Domain Adaptation of Pretrained Language Models: Case Studies on Biomedical NER and Covid-19 QA

Domain adaptation of Pretrained Language Models (PTLMs) is typically achieved\nby unsupervised pretraining on target-domain text. While successful, this\napproach is expensive in terms of hardware, runtime and CO_2 emissions. Here,\nwe propose a cheaper alternative: We train Word2Vec on target-domain text and\nalign the resulting word vectors with the wordpiece vectors of a general-domain\nPTLM. We evaluate on eight biomedical Named Entity Recognition (NER) tasks and\ncompare against the recently proposed BioBERT model. We cover over 60% of the\nBioBERT-BERT F1 delta, at 5% of BioBERT's CO_2 footprint and 2% of its cloud\ncompute cost. We also show how to quickly adapt an existing general-domain\nQuestion Answering (QA) model to an emerging domain: the Covid-19 pandemic.\n