Towards Student Learning Ability Estimation and Truth Discovery in Japanese Online Course

This paper focuses on an important task in online courses, which is to estimate student learning ability from students’ answering records. The challenge of this task is to automatically estimate the learning ability for students and infer the true answer for each question without any supervision. Most of the existing methods solve these challenges by designing an optimization objective function. However, these approaches ignore the characteristics of students from different groups. Intuitively, outstanding students always provide correct answers and should be assigned higher weights compared with ordinary students. Based on this intuition, this paper proposes a new optimization framework by dividing students into two groups: an authoritative group and an ordinary group. The losses of the objective are from both authoritative students and ordinary students. Through optimizing both losses simultaneously, the proposed model can automatically estimate learning ability for authoritative students and ordinary students and infer the right answer for each question. Two experiments conducted on two datasets show that the proposed model outperforms state-of-the-art baselines.