Research on Adaptive Cognitive Diagnostic Test Model based on Multilayer Bi-GRUs

Considering the deficiencies of the existing cognitive diagnostic models, this study proposed an extraction method of item semantic feature based on BERT. These features, together with Q-matrix, guessing parameter, slipping parameter and participants’ response results were integrated through concat layer and became feature matrix of response sequence. With the feature matrix as input, the cognitive diagnostic model was constructed based on N layers Bi-GRUs (Bi-directional Gated Recurrent Units) which transformed the feature matrix into feature vectors. Finally, non-linear classifier sigmod was adopted to classify each knowledge state dichotomously and loss function was designed to optimize the training effects of model. Training data from actual receptive tests, the model obtained 91.3% of model accuracy and 95.76% of average attribute accuracy. The experimental results show that the integration of item semantic features which embody lexical and syntax information of the items is beneficial for the model to predict participants’ language comprehension ability and provide individualized knowledge feedback. Besides, the deep bi-neural network model could comprehensively learn the deep information of participants’ response sequence and enhance the model effects of CD-CAT.