Search

Recent advances in Pre-trained Language Models (PrLMs) have driven general domain multi-choice Machine Reading Comprehension (MRC) to a new level. However, they perform much worse on domain-specific MRC like biomedicine, due to the lack of effective matching networks to capture the relationships among documents, question and candidate options. In this paper, we propose a Sketchy-Intensive (SkeIn) reading comprehension model, which simulates the cognitive thinking process of humans to solve the Chinese multi-choice biomedical MRC questions: (1) obtaining a general impression of content with a sketch reading process; (2) capturing dedicated information and relationships of documents, question and candidate options with an intensive reading process and making the final prediction. Experimental results show that our SkeIn model achieves substantial improvements over competing baseline PrLMs, with average accuracy improvements of +4.03% dev/+3.43% test, +2.69% dev/+3.22% test, and 5.31% dev/5.25% test from directly fine-tuning BERT-Base, BERT-wwm-ext and RoBERTa-wwm-ext-large, respectively, indicating the effectiveness of SkeIn to enhance the general performance of PrLMs on the biomedical MRC tasks.