Creation of Tonality Dictionaries using the Word Proximity Graph

The task of automatically analyzing the opinions of users about goods or services is extremely demanded. While choosing a product and planning a purchase most customers try to analyze the feedback of experiences users who already have tried it. Therefore, there is a need to create methods that determine the tonality of the text and the attributes of an item as precisely as possible. One of the methods for determining the tonality described in this paper is a semi-supervised algorithm using the word proximity graph. Different models of word embedding are compared: Word2Vec, FastTtext, model based on Singular Value Decomposition of the matrix of point-to-point mutual information between words (PPMI). The quality of the algorithm is evaluated on the corpora of reviews about the goods of the largest Russian consumer electronic retail chain M.Video. The work presents experiments that demonstrate how features based on word tonality can boost the quality of the prediction when classifying texts by tonality.