Computerized adaptive testing using neural networks

Abstract

Computerized Adaptive Testing (CAT) is a growing mode of assessment in many educational as well as non-educational institutions around the world. A distinct feature of CAT is its ability to tailor the test to the ability level of a test taker based on the observed responses to previously administered items. Hence, shorter tests can be administered and yet more accurate estimates can be attained. In the literature, several approaches have been proposed for making CAT systems based on Item Response Theory (IRT) and Bayesian Belief Networks (BBNs). Along with other detriments, these approaches rely on strict assumptions and require a large amount of probability data. In this thesis, we survey the state-of-the-art of computerized adaptive testing. Then, we explore the application of several static neural network models in designing traditional tests. After that, we propose a novel approach for making CAT systems using Recurrent Neural Networks (RNNs). As a form of neural networks, RNN is data-driven and self-adaptive classifier. Moreover, being recurrent, it captures the system dynamics by remembering the time-varying pattern of examinee's responses to previously administered items. We developed a toolkit of CAT system based on RNNs in Matlab 7.0 and evaluated its performance on several datasets. The results of the proposed approach are found to be promising as compared to some existing techniques.