The article presents a method that improves the quality of classification of objects described by a combination of known and unknown features. The method is based on modernized Informational Neurobayesian Approach with consideration of unknown features. The proposed method was developed and trained on 1500 text queries of Promobot users in Russian to classify them into 20 categories (classes). As a result, the use of the method allowed to completely solve the problem of misclassification for queries with combining known and unknown features of the model. The theoretical substantiation of the method is presented by the formulated and proved theorem On the Model with Limited Knowledge. It states, that in conditions of limited data, an equal number of equally unknown features of an object cannot have different significance for the classification problem.