Development of Methods for Automatic Testing of Sentiment Frames Compiled by Experts

There are two main approaches for sentiment analysis in the Natural Language Processing domain: rule-based approach and machine learning. Both of them use the various lexical resources such as thesaurus and sentiment frames as a base. However, there is still no research dedicated to examination of this type of manually annotated data. The aim of this study is to develop the algorithm in order to test the polarities of the words from the sentiment frames of RuSentiFrames lexicon. For this purpose, we consider the lexicon presented there by carrying out a number of experiments based on context distribution of words and semantic distance between them. During our study four approaches to determining the polarity of a word by context were examined, various combinations of sets considered as a gold standard and a test samples were tested, as well as three vector models and a number of other parameters. Our study provides a consistent approach for determining the polarity of words, as well as a list of recommendations for further research in this area.