Аннотация:The computational complexity of obtaining vector representations of words is constantly growing due to both an increase in the volume of input data and an increase in the complexity of the models. The design and training of a new model, which would take into account additional expert information about the semantic similarity between words, today is estimated at thousands of hours of computational experiments. In the study, existing vector representations are adjusted by transforming them into secondary representations taking into account the semantic similarity. The representation transformation and the learning procedure are proposed. Thus, instead of building a new model, the existing one is corrected, which significantly reduces computational costs compared to developing a new model.