The paper considers the generative-discriminative model (GAN) as applied to the task of analyzing text data, in particular, determining the gender of the author of a Russian-language text. The approach developed belongs to semi-supervised learning algorithms, when both labeled and unlabeled samples are involved in the model fitting process. The GAN model is implemented as a deep neural network consisting of fully connected, recurrent and convolution layers. The basis of the generative part of the GAN model is a variational auto-encoder, which encodes the input sample into the space of hidden variables and then decodes the latter into the original representation. When decoding, the class label of the input example is used, known in the case of the labeled set or predicted by the classifier for unlabeled samples. The input for the model is a sequence of words, each encoded by a vector of the principal components of its morphological features. To provide the function of recovering texts with more than 50 words, the principles of the work of language models are used. The discriminant part is configured to determine whether a given sample was generated by an auto-encoder or taken from the original set. Quality assessment of the GAN model was carried out on a set of texts from LiveJournal blogs. It is shown that the use of the generative-discriminative model allows to improve the quality of classification by 2 % in the F1 metric, while reducing the standard deviation by 2–3 times when learning on a small number of labeled examples. Various modes of training and variations in the topology of the GAN model are investigated, and the most effective modes of operation of models of this type for the task of classifying texts are demonstrated.