Abstract
In this paper we explore 5 similarity measures on the task of synonym extraction. We apply these measures to distributional word vectors created by word2vec model trained on a Russian-language corpus of user reviews. Through series of comprehensive experiments we have established that Cosine measure demonstrates, in general, the best performance in synonym extraction task. However, we found the way to improve Cosine measure performance correcting its value with the help of Gower distance. Our new corrective measure which we call Cosine-Gower, outperforms Cosine similarity for all meaningful parts of speech excluding verbs.