Remember the Facts? Investigating Answer-Aware Neural Question Generation for Text Comprehension

Reading is a crucial skill in the 21st century. Thus, scaffolding text comprehension by automatically generated questions may greatly profit learners. Yet, the state-of-the-art methods for automatic question generation, answer-aware neural question generators (NQGs), are rarely seen in the educational domain. Hence, we investigate the quality of questions generated by a novel approach comprising an answer-aware NQG and two novel answer candidate selection strategies based on semantic graph matching. In median, the approach generates clear, answerable and useful factual questions outperforming an answer-unaware NQG on educational datasets as shown by automatic and human evaluation. Furthermore, we analyze the types of questions generated, showing that the question types differ across answer selection strategies yet remain factual.