Treffer: Automatic Question Generation to Support Reading Comprehension of Learners - Content Selection, Neural Question Generation, and Educational Evaluation

Simply reading texts passively without actively engaging with their content is suboptimal for text comprehension since learners may miss crucial concepts or misunderstand essential ideas. In contrast, engaging learners actively by asking questions fosters text comprehension. However, educational resources frequently lack questions. Textbooks often contain only a few at the end of a chapter, and informal learning resources such as Wikipedia lack them entirely. Thus, in this thesis, we study to what extent questions about educational science texts can be automatically generated, tackling two research questions. The first question concerns selecting learning-relevant passages to guide the generation process. The second question investigates the generated questions' potential effects and applicability in reading comprehension scenarios. Our first contribution improves the understanding of neural question generation's quality in education. We find that the generators' high linguistic quality transfers to educational texts but that they require guidance by educational content selection. In consequence, we study multiple educational context and answer selection mechanisms. In our second contribution, we propose novel context selection approaches which target question-worthy sentences in texts. In contrast to previous works, our context selectors are guided by educational theory. The proposed methods perform competitive to related work while operating with educationally motivated decision criteria that are easier to understand for educational experts. The third contribution addresses answer selection methods to guide neural question generation with expected answers. Our experiments highlight the need for educational corpora for the task. Models trained on noneducational corpora do not transfer well to the educational domain. Given this discrepancy, we propose a novel corpus construction approach. It automatically derives educational answer selection corpora from textbooks.