Introduction: diagnosing mental illness is a complex process that includes conducting dialogue conversations, analyzing the behavior of the subject and passing specialized tests. The successful solution of this problem can be influenced by both the lack of knowledge and experience of the psychologist, and the presence of contradictory or incomplete initial data on the part of the patient. To eliminate this drawback, expert-based or intelligent systems are being developed. Purpose: development of a technique for determining the mental state of social network users. Results: using machine learning methods, a technique has been developed designed to determine the type of a mental state of social network users. The novelty of the proposed technique is in the usage of a two-step text preprocessing procedure and the construction of several sets of features which describe the emotional mood of social network users at the level of the messages published by them. As the initial data, we have used text messages of users of the social network Reddit. There are three stages in the technique: 1) data collection, 2) data preprocessing, 3) post labeling and feature construction. To assess the functioning of a software tool built on the basis of this technique, four indicators were used: accuracy, precision, recall, and F-measure. The best results are demonstrated with a One-vs-Rest ensemble using linear support vector machines as basic solvers. Practical relevance: the investigation results can be used in the construction of auxiliary systems that are aimed at supporting decision-making by psychologists in determining mental disorders.