Predicting Prerequisite Relations for Unseen Concepts

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8542 - 8548 December 7-11, 2022 ©2022 Association for Computational Linguistics Predicting Prerequisite Relations for Unseen Concepts Yaxin Zhu and Hamed Zamani Center for Intelligent Information Retrieval University of Massachusetts Amherst {yaxinzhu, zamani}@cs.umass.edu Abstract Concept prerequisite learning (CPL) plays a key role in developing technologies that assist people to learn a new complex topic or con- cept. Previous work commonly assumes that all concepts are given at training time and solely focuses on predicting the unseen prerequisite relationships between them. However, many real-world scenarios deal with concepts that are left undiscovered at training time, which is relatively unexplored. This paper studies this problem and proposes a novel alternating knowledge distillation approach to take advan- tage of both content- and graph-based models for this task. Extensive experiments on three public benchmarks demonstrate up to 10% im- provements in terms of F1 score. 1 Introduction As the amount of online educational data rapidly grows, it is more important than ever to develop information access systems that assist users to learn new complex topics or concepts (Gwizdka et al., 2016; Collins-Thompson et al., 2017; Eickhoff et al., 2017; Urgo and Arguello, 2022). A fun- damental step towards developing these systems is concept prerequisite learning (CPL)—the task of building a concept graph by structuring open knowledge in prerequisite relations. A prerequisite is a directed relation between two concepts, e.g., Heap Tree is a prerequisite of Heap Sort. CPL was first introduced by Talukdar and Cohen (2012) with the aim of formulating probabilistic planning problems for machine learning solvers. A number of CPL models use various kinds external resources including Wikipedia links (Liang et al., 2015), textbook structures (Wang et al., 2016), and course dependencies (Liang et al., 2017; Liu et al., 2016) in two ways: content-based (Pan et al., 2017; Roy et al., 2019; Gasparetti, 2022) and graph-based (Liang et al., 2015). There also exists research on active learning (Liang et al., 2018a,b), unsuper- vised learning (Li et al., 2020), and domain adap- tation (Li et al., 2021) to meet data insufficiency challenges in CPL problems. However, most exist- ing approaches assume the system to reconstruct concept prerequisite paths with vague knowledge (i.e. incomplete relations) of each concept, or to transfer graph structure information to a new do- main. In practical scenarios, knowledge is updated with both new concepts and relations introduced. To jump out of the offline setting of graph comple- tion with given concepts, we define a new task - CPL for unseen concepts, i.e. predicting prerequi- site relationships for concepts that never appear in the training set. Most existing graph-based CPL approaches can- not be simply used for unseen concepts, because the randomly initialized concept representations do not get updated for unseen concepts. A simple solu- tion would be initializing the concept embeddings based on content-based models (Li et al., 2019; Jia et al., 2021; Zhang et al., 2022; Sun et al., 2022). To better take advantage of …

Content truncated to 500 words