The paper presents the results of applying machine learning algorithms to the task of automatic evaluation of verbal and noun collocations. The study of collocability showed that distribution models can be successfully used to model relations within phrases. A phrase is considered to be significant if its vector representation is close to the vector representation of the headword. We used the following methods for evaluating collocations based on machine learning and word embeddings: baseline, the method of analogy and linear transformation. Automatically selected phrases were compared with the data provided in lexicographic sources (in explanatory dictionaries and collocation dictionaries, five resources were considered in total), which formed the so-called gold standard. The results showed that the methods under consideration are successfully used to extract phrases, including those that are not reflected in the dictionaries. These examples can gain lexicographers’ attention , although they are not given in the resources and need additional expert evaluation. Therefore, it is necessary to further compare the algorithms with other statistical metrics and increase the number of phrases in the gold standard.