In this paper we discuss the task of prepositional phrase classification in the Russian annotated corpus of prepositional phrases. As previous research has shown, differentiation of highly confused classes, namely THEME and OBJECT classes, remains a problem waiting for the computational solution. Since simple classifier architecture demonstrates significant performance on these classes, we propose a tree-based classifier architecture to improve performance on the whole and these classes specifically. This architecture consists of a main classifier validating its decisions concerning troublesome classes with another supporting classifier, trained to differentiate between the classes causing performance dropdown. We experiment with various types of classifiers inside of our architecture and various embedding models for the Russian language, which we use for encoding the dataset. The best result that we managed to achieve is an overall F1-score of 0.76 on the validation set using the classifier trained with DeepPavlov/rubert-base-cased model and SVM (Support Vector Machines) classifiers.