Combining Wikipedia to Identify Prerequisite Relations of Concepts in MOOCs

Abstract

Many applications like the personalization recommendation system of online learning are based on prerequisite relations of concepts, which prompted us to automatically infer the prerequisite relations between the concepts in Massive Open Online Courses (MOOCs). The previous methods mostly use artificial features to identify the prerequisite relations from learning materials and Wikipedia. However, artificial features are complicated to deeply mine prerequisite information in MOOC videos and the Wikipedia-directed graph, resulting in poor performance. We propose a new and more effective method to identify prerequisite relations from the above two data resources. We first use a graph embedding algorithm to learn the vector representations of concepts from the created Wikipedia-directed graph and use the cosine similarity between the vectors to represent the semantic and structural relevance between the concepts. Second, we pre-train a Siamese network whose inputs are representations of course concepts learned by a variation of the LDA model to find more practical information of prerequisite relations from MOOC subtitles. Then, the concept similarities related to topic distribution can be represented by the pre-trained Siamese network's outputs. Finally, we add some excellent artificial features to expand the information of the prerequisite relations and input them together into a binary classifier to identify the prerequisite relations of the concepts in MOOCs. Our experiments on two MOOC datasets indicate that the proposed method achieves significant improvements comparing with existing methods.