Methods of lexicon integration with machine learning for sentiment analysis system

This paper describes the application of SVM classifier for sentiment classification of Russian Twitter messages in the banking and telecommunications domains of SentiRuEval-2016 competition. Varieties of features were implemented to improve the quality of message classification, especially sentiment score features based on a set of sentiment lexicons. We study the impact of different training types (balanced/imbalanced) and its volumes, and advantages of applying several lexicon-based features. Before SentiRuEval-2016, the classifier was tuned on the previous year collection of the same competition (SentiRuEval-2015) to obtain a better settings set. The created system achieved the third place at SentiRuEval-2016 in both tasks. The experiments performed after the SentiRuEval-2016 evaluation allowed us to improve our results by searching for a better ’Cost’ parameter value of SVM classifier and extracting more information from lexicons into new features. The final classifier achieved results close to the top results of the competition.