Autogenerating a Domain-Specific Question-Answering Data Set to Enable High-Performing Language Models for Magnetic Materials
This paper presents a method for autogenerating domain-specific question-answering data, builds MagQA with 168,080 magnetic-materials QA pairs, and uses it to finetune BERT-style models; on a manually annotated magnetic-materials test set, a vanilla BERT-base-cased finetuned on MagQA mixed with SQuAD v2 (MagBERT_MagQA_Mixed) achieves the best result with an F1 of 78.43% and an exact-match score of 72.84%, indicating that, given sufficiently large and high-quality domain-specific QA data, domain-specific BERT models need only be finetuned from vanilla BERT without domain-adaptive pretraining.