Abstract
Pre-trained language models have brought significant performance improvements in many natural language understanding tasks. Domain-adaptive language models, which are trained with a specific domain corpus, exhibit high performance in their target domains. However, pre-training these models with a large amount of domain-specific data requires a substantial computational budget and resources, necessitating the development of efficient pre-training methods. In this paper, we propose a novel subset selection method called AlignSet, which extracts an informative subset from a given domain dataset for efficient pre-training. Our goal is to extract an informative subset that enables faster learning of the language model compared to learning from the entire dataset. By experiments across multiple domains, we demonstrate that AlignSet generates better subsets than other methods.