G-DINA认知诊断模型在语言测验中的验证

Validating G-DINA Model in Language Test Diagnosis

G-DINA模型是DINA模型的一般化模型,具有补偿性和饱和性两个主要特征。此项研究以代表性的语言测验类型阅读测验为案例,应用G-DINA模型对1029名被试的PISA英语阅读测验结果进行实证分析,发现G-DINA模型不仅能准确诊断被试对各阅读技能的认知状态,还能揭示阅读技能之间的关系。研究证明了两个假设:补偿饱和型认知诊断模型对多元抽象的语言测验有较高的适应性;G-DINA这一新生认知诊断模型可以被用来诊断较为复杂抽象的语言测验,且经得起统计学和语言学理论的双重考验。

Since language tests are characteristic of multidimensionality and abstract skills, cognitive diagnosis on language tests is not only a challenge to research about cognitive diagnostic assessments but also important evidence to validate cognitive diagnostic models. Cognitive diagnostic models like Rules Space Methodology, Attribute Hierarchy Method, Fusion Model, and General Diagnostic Model have been applied to various types of language tests. However, all those cognitive diagnostic models are applied in noncompensatory or reduced ways. The G - DINA model, a generalized model of DINA model, is a compensatory and saturated cognitive diagnostic model. Under a compensatory cognitive diagnostic model, successfully executing only a few or some of the attributes required for an item can contribute to the probability of correct response to that item. A saturated cognitive diagnostic model not only includes all single - skill at- tributes required by items but also takes all possible attribute interactions as mixed - skill attributes. The compensatory nature of the G - DINA model caters to the multidimensionality of language tests ; the saturated nature of the G - DINA model caters to the abstractness and indivisibility of language skills. Taking the reading test, a typical form of language tests, as the case study, this paper applies the G - DINA model to the empirical analysis of the PISA English reading test results of 1029 British subjects. The study aims to verify two hypotheses ： compensatory and saturated cognitive diagnostic models can ideally fit language tests ; the G - DINA model, a new cognitive diagnostic model, can be applied to diagnose complicated and abstract language tests. Through substantive analysis of item content by 6 experts, the 5 reading attributes and the Q - matrix are defined and constructed. The data are analyzed with the G - DINA framework code operated with the OxEdit software. The absolute model fit of the analysis is based on the residual between the observed and predic- ted correlation of item pair with the Fisher transformation （p） and the residual between the observed and predicted log -odds ratios （LOR） of pair - wise item responses （l） jointly. The absolute model fit of this study reaches a very high level where the p value under the hypothesis （residuals = 0） is over. 10. Three pieces of further evidence are discovered to support the hypotheses： （1）the mastery probability of linguistically lower - level attributes is higher than that of linguistically higher - level attributes, which demonstrates that the compensatory nature of the G - DINA model caters to the muhidimensionality of language tests ; （2） the compensatory and saturated nature of the G - DINA model can be reflected in revealing the relationships among the reading attributes which are interrelated to and interdependent on one another; and（3）the compensatory and saturated G- DINA model can detect linguistically explainable item problems, which demonstrates the accuracy of diagnosis with that model.