Can Machines Learn to Comprehend Scientific Literature?

Abstract

To measure the ability of a machine to understand professional-level scientific articles, we construct a scientific question answering task called PaperQA. The PaperQA task is based on more than 80 000 "fill-in-the-blank" type questions on articles from reputed scientific journals such as Nature and Science. We perform fine-grained linguistic analysis and evaluation to compare PaperQA and other conventional question and answering (QA) tasks on general literature (e.g., books, news articles, and Wikipedia texts). The results indicate that the PaperQA task is the most difficult QA task for both humans (lay people) and machines (deep-learning models). Moreover, humans generally outperform machines in conventional QA tasks, but we found that advanced deep-learning models outperform humans by 3%-13% on average in the PaperQA task. The PaperQA dataset used in this paper is publicly available at http://dmis.korea.ac.kr/downloads?id=PaperQA.