The task of information retrieval is to find information that best satisfies the user's information needs. In the context of social media, information retrieval is complicated by the high dynamism of content, thematic heterogeneity, and the diversity of users' mental models. This paper proposes an approach to solving the problem of information retrieval under such conditions by constructing a multi-domain dynamic knowledge system. Its novelty lies in the combination of three levels of semantics: problem-oriented, represented by the ontology of the metatask (describing the search objectives); domain-specific, implemented through a dynamic multi-layer knowledge graph built on the basis of user content of social media; and domain-independent, based on a lexical database and a large language model. The knowledge graph allows us to reflect various contexts of concept usage corresponding to thematic clusters in the document collection. Such integration allows us to take into account the evolution of concepts, discourse features, and mental stereotypes of communication participants. To evaluate the effectiveness of the proposed system, an experiment was conducted using a dataset of publications from the VKontakte social network for problem-oriented monitoring of publications, where the selection of relevant publications from non-thematic sources is required. To solve this problem, a technology based on the use of the distance metric between query terms and publication terms in a multi-layer knowledge graph was proposed. The results of the experiment using this technology confirm the effectiveness of the proposed model for information retrieval tasks compared to standard keyword search and embedding models. In continuation of this study, it is planned to create a lexical database and also to consider the possibility of expanding the model by using a measure of pointwise mutual information and graph embedding methods.