Due to the development of computer network, a large amount of documents are treated in many fields. The number of digital document data stored in databases is enormous, accordingly it is difficult for analysts to read all documents and classify it by hand. Therefore, it is necessary to develop the technology of automatic document classification by using computers these days. From the above needs, many classifiers with good performance have been proposed, i.e., Relevance Vector Machine (RVM) and Support Vector Machine (SVM) that are known as good binary classifiers. For multi-valued document classification problems, it is known that a multi-valued classifier by combining several binary classifiers has a good performance. In this study, the method to construct an efficient combination of binary classifiers based on improving Generalized Bradley-Terry (GBT) model, which has high extensibility, is focused. This model is an expansion of Bradley-Terry (BT) model. Though the BT model has a limitation on combination of classes, the GBT model enables us to utilize any binary classifier which classifies into two arbitrary subsets in the class set. Generally, when several binary classifiers learn from the training dataset, there would be the difference of accuracy between these binary classifiers, due to the existence of categories that cannot be easily classified. However, the conventional method of multi-valued classification by GBT binary classifiers does not take the accuracy of each classifier into consideration. To avoid this problem, a new way of multi-valued classification method by considering each classifier's accuracy is proposed. The purpose of this study is to construct a good multi-valued classifier by calculating the accuracy of each classifier and utilizing it as the weight. In order to verify the effectiveness of the proposed method, the simulation experiment by using newspaper articles is conducted.