SEEP: Semantic-enhanced question embeddings pre-training for improving knowledge tracing

Knowledge Tracing (KT) defines the task of diagnosing students’ dynamic knowledge level in exercises. Although existing efforts have leveraged question information, most of them either learn question embeddings during the process of model training or represent questions based on the correlation between questions and concepts, which ignores plentiful implicit information entailed in the student-question-concept interaction and the revelation of fine-grained semantics between the interaction as well as the usage of the students’ historical answers. It is, however, challenging to extract and refine implicit information in the student-question-concept interaction which is highly heterogeneous and complex. To this end, in this paper, we present a novel Semantic-Enhanced Question Embeddings Pre-training (SEEP) method, concentrating on decomposing underlying relation information in the interaction and further fusing information of questions and concepts under different decomposed semantic perspectives to obtain semantic-enhanced question embeddings for improving performances of KT methods. Extensive experiments conducted on two real-world datasets show SEEP has the higher expressive power that enables KT methods to predict students’ performance.