This paper considers the problem of automating the previously presented scheme of classification of users’ posts in social media. Different variations of the embeddings obtained from the RuBERT and ELMo language models are used as input features of the classification. Theoretical significance lies in the adaptation of existing language models BERT and ELMo in order to improve the accuracy of solving the applied problem of posts classification. The practical significance is determined by further automation of the classification of users’ text posts, which will form the basis of the system for assessing the expression of their personality traits and, indirectly, vulnerabilities to social engineering attack.