如何通过测验了解学生的知识掌握程度，测验项目编制是否合理是教育工作者非常重要的研究课题. 认知诊断理论可以说为此课题的研究打开了一面崭新的窗户，是认知心理学与现代测量学相结合的产物，是目前国内外心理测量学研究的热点. 目前已提出的重要方法有规则空间模型(RSM)，属性层级方法(AHM)，DINA模型，G-DINA模型等，这些方法能够较好地处理0-1评分问题，但对于实际更复杂的多级评分问题就不适合了. 近几年部分学者提出了几种用于多级评分的模型，但仍在发展之中；另外，对于0-1评分以及多级评分中的项目相关性计算与分析仍是一个空白.本文针对多级评分问题基于现有模型提出了新的改进模型. 首先本文提出项目相关性的概念，并把项目之间的相关信息引入DINA模型，建立将多级评分问题化简为0-1评分问题 E-DINA模型，为解决多级评分问题提供一种有效途径；其次基于G-DINA模型的思想，引入滑动矩阵的概念给出一种具有普适性的多级评分项目的认知诊断方法，即GPCDM；第三是将0-1评分的广义距离法推广到多级评分情形，建立P-GDD模型，该模型既可用于属性相等权重的多级评分项目，也可用于属性不等权重的多级评分项目.针对这些新提出的模型，本文给出了大量的模拟计算以验证其有效性，结果表明：（1）E-DINA模型可以很好地分析项目相关性问题，通过和DINA模型的对比表明，该模型有效地提高了模式和属性判准率，通过将复杂的多级评分问题化简为0-1评分问题，大大地扩展了DINA模型的应用范围.（2）GPCDM适用范围较广，具有普适性，只要评分标准能给出知识状态到理想反应模式的对应关系，就可以用该模型估计被试的知识状态；该模型需要有更多的被试人数才能有效地估计出这些参数，也因此得到了更高的判准率.（3）P-GDD作为一种基于IRT的诊断方法继承了IRT的优点，模拟结果也表明该方法具有较高的模式和属性判准率. 在分析实际数据时，该方法简单易用，同时还具有将属性掌握的程度分成不同等级的优点.总之，根据多级评分问题的实际特点或评分准则，有针对性地选择使用本文提出的三种模型，可以为被试者和教育工作者提供更多有价值的信息.

It is essential for educators to figure out whether and how they can evaluate students' mastery of knowledge through a test, as well as to examine if the test items can achieve the goal. Cognitive diagnosis theory, which is a combination of psychology and modern psychometrics, has opened a new window for the research of this subject, and it is a hot topic in the area of psychometric research domestic and abroad. So far, some important methods proposed are: Rule Space Model (RSM), Attribute Hierarchy Model (AHM), DINA Model, Generalized-DINA Model. These methods are good at dealing with dichotomized (0-1) problems, while they are inappropriate for polytomous problems, which is much more complicated than 0-1 problems.Recent years, some scholars have raised solutions to the polytomous problems, and these new methods are still in development. Furthermore, it still is a blank for the research on the problems with correlations between items, which are either 0-1 problems or polytomous problems.In this article, some improved models are proposed based on existed cognitive diagnosis models to deal with polytomous problems. First the correlation between items is proposed, and is introduced into DINA model, then, the polytomous problems can be easily converted into dichotomize problems, the improved model is called E-DINA model. Second, a generalized cognitive diagnosis method, called GPCDM, for polytomous problems is proposed through introducing sliding matrix concept. Finally, the Generalized Distance Discrimination method used for 0-1 problems is extended to polytomous cases.The article gives quantities of simulations to test the validity of these new models. The results have shown, (1) E-DINA model can be used to analyze items correlations very well, and achieves higher Pattern Match Ratio and Attribute Match Ratio through comparison with DINA model. This results in that any polytomous problems can be simplified to dichotomized problems and analyzed conveniently and easily by using E-DINA model. Thus, DINA model is extended to more widespread problems. (2) GPCDM has generality and can be used to a wider range of applications. It can provide examinees' knowledge statements if the correspondences between knowledge statements and expectation response patterns are known in the rating criteria. In order to effectively estimate related parameters, more examinees are needed, and the benifit is higher Pattern Match Ratio and Attribute Match Ratio to be achieved. (3) P-GDD, a diagnosis method based on IRT, inherited IRT advantages. The simulation results also prove that P-GDD can achieve higher Pattern Match Ratio and Attribute Match Ratio. This method is simple and can be used easily for real data analysis, meanwhile it has one more advantage to be able to evaluate examinee' mastery of knowledge in different levels. In summary, to properly select the models presented in the article according to the properties in real polytomous problems or rating criteria, can provide examinees and educators more valuable information.