LTWNN: A Novel Approach Using Sentence Embeddings for Extracting Diverse Concepts in MOOCs

Abstract

AbstractAs a global online education platform, Massive Open Online Courses (MOOCs) provide high-quality learning content. It is a challenging issue to design a key course concept for students with different backgrounds. Even though much work concerned with course concept extraction in MOOC has been done, those related works simply utilize external knowledge to get the relatedness of two different candidate concepts. Furthermore, they require the input to belong to multi-document and severely rely on seed sets, in which their model shows poor performance when input is a single document. Addressing these drawbacks, we tackle concept extraction from a single document using LTWNN, a novel method Learning to Weight with Neural Network for Course Concept Extraction in MOOCs. With LTWNN, we make full use of external knowledge via making relatedness between each candidate concept and document by introducing an embedding-based maximal marginal relevance (MMR), which explicitly increases diversity among selected concepts. Moreover, we combine the inner statistical information and external knowledge, in which the neural network automatically learns to allocate weight for them. Experiments on different course corpus show that our method outperforms alternative methods.KeywordsMOOCCourse concept extractionLearning to weightNeural networkDiversity