This paper presents the development of a recurrent neural network (RNN) architecture for the classification of reviews of the Internet Movie Database (IMDB) dataset. A number of experiments were conducted for various RNN models, namely SimpleRNN, LSTM, and GRU. As a result of the study, the optimal LSTM architecture with 32 neurons was determined. This model achieved accuracy for training and test data of 0.95 and 0.88, respectively. In comparison, the GRU model with similar parameters achieved accuracy for test data of 0.87 with a training time 2.1 times less than the LSTM model. Therefore, this model is suitable for tasks where the time of the task is important. Further experiments were conducted to explore different configurations of the models, including varying the number of training epochs, neurons, and recurrent layers. However, none of these configurations surpassed the performance achieved by the LSTM with 32 neurons in terms of accuracy. The results obtained in the study provide useful information about reviews of specific films and can become the basis for a system of recommendations for watching films taking into account the interests and preferences of users.