Abstract
Vector representation of words plays a major role in the natural language processing. As a fundamental unit, the vector representation is further used to solve applied problems: text classification, textual entailment, named-entity recognition. In recent times was created several models to produce word embedding vectors which uses different approaches. However, they suffer from accidents while training, such as random initialization of weights, the random order of the examples. Therefore, it is impossible to reproduce the result, and repeated experiments using the same dataset and algorithms lead to various close results. In this work, we presented methods for estimating the “dissimilarity” of the semantic spaces built by the algorithms of word embeddings and give mathematical intuition about influence of various randomness on the structure of semantic spaces.