We propose a novel approach to compute a model of semantic structure of natural language texts corpora based on the models of the texts. This approach di ers from LSI and standard ways of topic modeling (D. Blei, 2012). We do not use matrix factorization and probabilistic distributions of the words given a priory. As the base for the corpora model construction we use the set of the original vector models of the corpora texts. The ways of empirical validation and applications of the models are considered.