Abstract

Knowledge tracing (KT) aims to trace a learner’s understanding or achievement of knowledge based on learning history. The surge in online learning systems has intensified the necessity for automated measurement of students’ knowledge states. In particular, in the case of learning in the English proficiency assessment field, such as TOEIC, it is required to model the knowledge states by reflecting on the difficulty of questions. However, previous KT approaches often overly complexify their model structures solely to accommodate difficulty or consider it only for a secondary purpose such as data augmentation, hindering the adaptability of potent and general-purpose models such as Transformers to other cognitive components. Addressing this, we investigate the integration of question difficulty within KT with a potent general-purpose model for application in English proficiency assessment. We conducted empirical studies with three approaches to embed difficulty effectively: (i) reconstructing input features by incorporating difficulty, (ii) predicting difficulty with a multi-task learning objective, and (iii) enhancing the model’s output representations from (i) and (ii). Experiments validate that direct inclusion of difficulty in input features, paired with enriched output representations, consistently amplifies KT performance, underscoring the significance of holistic consideration of difficulty in the KT domain.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call