Research Of IRT-based Computerized Adaptive Testing System

Computerized Adaptive Testing is the fully integrated product of modern educational and psychological measurement theory and computer technology. CAT gets more and more attention and application with the rapid development in computer technology. However, the limitations of the traditional selection strategies cause a series of problems, such as test security, item exposure rate, test efficiency, item bank utilization and so on. Although many scholars have done a great deal of research on these issues and got some effective solutions, most of these new solutions are stay in the simulation stage and not practical, and also these studies are independent of each other, not effectively integrated to solve practical problems.To solve these problems, this paper first analyses the theory of computer adaptive testing and its development, mainly studies the Item Response Theory and three commonly used logistic models, and compares the various parameters on the impact of adaptive testing. Then, deeply analyses the four key parts of CAT, item bank, administering the test, capacity estimation and termination rule. And the selection strategies have been specially studied:Through the analysis of three relatively mature adaptive testings, Stratified method, Sympson-Hetter method and Modified Multinomial Model, an improved selection strategy CSTR_MMM_SH is proposed. This strategy combines Stratified method, Sympson-Hetter method and Modified Multinomial Model. First, it is based on stratification, stratified the item bank (the content coverage of each stratum is similar to that of the full item pool; the distribution of the b-parameters in each stratum is as similar as possible to that of the full item pool; and the average value of the a-parameters increases across strata), and the corresponding tests are also divided into several stages. So it exhibits the advantage of Stratified method that administers items with relatively lower discrimination in the early stages of a CAT, and administers items with relatively higher discrimination during later stages. Then, make up the shortage of Stratified method at item exposure control by SH, to get the more balanced distribution of exposure. Finally, MMM method is used to balance the contents of the entire test, so makes the testing process more reasonable and the test results more comparable. Additionally, based on this improved selection strategy, this paper designs an adaptive testing system. The syetem uses three-parameter logistic model to design item bank, maximum likelihood method to estimate the abilities of examinees, CSTR_MMM_SH selection strategy to sleet items, combination of fixed-length and no fixed-length method to end the test.