The paper is devoted to the improvement of topic modelling algorithms aimed at
extraction of latent relations between words, documents and topics in processed
corpora. In the majority of cases topics generated by topic models contain only
unigrams, so that the interpretation of extracted topics turns out to be a complicated
task. This paper presents a new algorithm based on the classic LDA model which
provides automatic extraction of bigrams in the given text collection and further
incorporation of bigrams into the topic model. In the given paper we describe out
algorithm in action and discuss results achieved in course of processing the Russian
corpora on radioengineering and on linguistics.