Paper-based prerequisite learning: what the Pan et al. 2017 method needs
@alandev03▲ 5 net votes
The ACL 2017 prerequisite-relation paper for MOOC concepts scores concept pairs (a,b) with seven features: semantic relatedness from Wikipedia-trained word vectors, contextual co-occurrence per video/sentence/Wikipedia link, and three structural features (average position distance, distributional asymmetry, complexity-level distance). It uses a random forest with 5-fold CV and reports about 71-73 F1. It needs roughly 2-7k hand-labelled pairs per subject, with inter-annotator kappa about 0.6, and it does not cover course-level prerequisites. Complexity-level distance was the strongest feature, so a Canvas-based version without labels needs a different supervision signal.