What Is Relevant for Learning? Approximating Readers’ Intuition Using Neural Content Selection

Abstract

Experienced readers intuitively mark text passages containing central concepts as learning-relevant when reading actively. Although this intuitive process of marking important information is sometimes imperfect, it fosters comprehension. It would be beneficial to approximate this intuition by automatically detecting potential learning-relevant content. It is a building block for various upstream tasks such as automatic self-assessment or intelligent author assistance. This work argues that learners often apply heuristics based on different sentence types to determine the learning-relevant contents in texts. We show that such heuristics can be approximated using neural sentence classifiers and implement two neural sentence classifiers detecting causal and definitory sentences. We evaluate the classifiers’ ability to detect learning-relevant information in an empirical study (N = 37). Furthermore, a system performance evaluation compares the proposed classifiers with unsupervised summarization systems. We find evidence for a small but reliable association between the chosen automatically detectable sentence types (definition/causal) and the learners’ perception of content relevance. Additionally, the classifiers outperform most other relevant content selection techniques in our experiments. Interestingly, other simple heuristics based on sentence position or length also exhibit strong performance.