The article explores the capabilities of various models of recurrent neural networks in text data generation. Specifically, classic recurrent network (RNN), long short-term memory network (LSTM) and generative adversarial network (GAN) models are considered in the context of the problem of generating specification text for assembly drawings according to the format approved by government standard. To train the models, a data set in Russian language was used, expanded with additional records simulating input data, consisting of the drawing parts, and expected text of the specifications. The input data set for the study was divided into four groups of equal size, depending on three main factors: amount of input parts, their repeatability and grammatical complexity. It is concluded that for all four groups of input data generative adversarial networks (GANs) have the maximum ratio of error-free responses to all responses generated by model, followed by LSTMs and, lastly – RNNs. As a result, it is planned to use GAN-based models in future researches on specification text data generation.