Processing Social Media Posts on Mental Health Issues

Natural language processing methods work effectively with large amounts of data and are well suited for research on current topics in society. One of these topics is the presence of psychological problems in a person. In this study, we study methods for detecting such a psychological problem as depression using the analysis of social networks. The use of social media data provides a representative picture of the assessment of the real number of such problems, as people more sincerely and in detail talk about their psychological health and possible problems due to partial anonymity in social networks, so it is possible to detect the disease in time and provide the necessary assistance. When solving the basic problem of detecting the presence of depression, with the existence of a strong dissimilarity of positive and negative classes, the quality of F1 = 0.9728 was achieved. When moving on to the more complex task of identifying among all the posts mentioning depression in different contexts, posts that could with a high probability signal the presence of depression in its author, the quality of F1 = 0.7419 was achieved. It was shown that the use of popular methods of sentimental analysis for similar classification problems does not give the proper quality to our problem. A set of marked-up Russian-language data was also created (33% of the positive class), containing references to depression in each post in various contexts, there is no analogue to such marked-up data in any language. In addition to the above, various signs and characteristics of the manifestation of depression of different severity were obtained, which allows for the necessary interventions of different intensity and speed.