Assessment modeling frames prediction over student interaction sequences as a pre-training task. The model predicts an assessment feature after masking information that should not be visible for that prediction.

Google

Assessment modeling is a class of pre-training tasks for interactive educational systems. It predicts a masked assessment feature from the remaining interactions in a student's interaction sequence. The approach described in Assessment Modeling: Fundamental Pre-training Tasks for Interactive Educational Systems uses a deep bidirectional transformer encoder.

Assessment Modeling in Interactive Educational Systems

In a knowledge tracing task, there is a sequence of interactions for each student: $$[I_1, \cdots, I_T]$$, and each of these interactions consists of several features: $$\{f_t^1, \cdots, f_t^n\}$$. In knowledge tracing, the goal is to predict feature $$f$$ at time $$t$$, which is formulated as: $$p(f_t^k | \{ I_1, \cdots, I_{t-1}, I_t, I_{t+1}, \cdots, I_T \} \setminus \Gamma(f_t^k))$$ where $$\Gamma(f)$$ is a set of features that should be masked. While knowledge tracing typically predicts the correctness of a response, assessment modeling aims to predict the assessment given the sequence of interactions: $$p(a_t^k | \{ I_1, \cdots, I_{t-1}, I_t, I_{t+1}, \cdots, I_T \} \setminus \Gamma(a_t^k))$$

Learn Before

Related