Efficient Transformer-based Knowledge Tracing for a Personalized Language Education Application

The purpose of this paper is to propose a new deep learning-based approach to recommend highly personalized educational contents to learners. Towards this goal, we present a knowledge tracing algorithm by adding long short-term memory units to a Transformer-based model. By inferring the knowledge state of a learner through the proposed KT algorithm, it not only removes problems that the learner does not have to solve but also suggest problems so that the learner's knowledge state level improves most efficiently. In this manner, a personalized educational curriculum can be provided to each learner. We trained the model with 90 million datasets collected from a Hangeul (i.e., Korean character) learning mobile application called "Sojung-Hangeul", one of the market-leading Korean learning services. The experimental results show that the AUC of the proposed model significantly improves from 0.88 to 0.92 compared to the recent Transformer-based approach in real-time environments. The proposed deep learning model is applied into "Sojung-Hangeul", and the application is currently available at https://bit.ly/Sojung-Hangeul.