Resumen
Online social networks (OSN) are actively used for implementation of illegal and destructive activities including propaganda of drugs, suicide, terrorism. There is the task of detecting such materials for providing information-psychological security of a person. For dealing with it automated monitoring of OSN is required. Monitoring includes search in unstructured users? text messages using keywords characterizing the object of monitoring. There is a problem of thematic relevance during process of OSN monitoring as well as in information retrieval systems. It happens because of language phenomenon of lexical ambiguity. The task of illegal content detecting is complicated by using of slang and jargon in communications, it does not allow to use existing effective approaches to word sense disambiguation. For fixing the problem of topical relevance author suggests to use topic models based on contexts of keywords. Additional multidimensional scaling for contexts in semantic space and subsequent clustering allow to make sense induction of posts from OSN. Condition for classifying a message as illegal content is proposed. Developed technique was tested on The National Corpus of Russian and The General Internet-Corpus of Russian.