Assessing the Impact of Text Preprocessing in Sentiment Analysis of Short Social Network Messages in the Russian Language

Presently, social networks provide individuals with a unique ability to share personal opinions with others across the world. On one of such mediums, Twitter, personal opinions and general information are presented in the form of short texts (with a limit of 280 characters). Machine learning algorithms could be leveraged in detecting sentiments and opinions within such textual data. For more efficient results, text is usually transformed into a better state through the process of pre-processing. To detect the impact of this step, we used a corpus of Twitter posts in Russian language labelled either as positive or negative sentiments and applied various combinations of preprocessing techniques. Furthermore, we classified the transformed texts with a Naive Bayes Classifier and Logistic Regression and furthermore evaluated the accuracy. Results for each combination of preprocessing type and model were compiled into tables and the conclusions on impact of preprocessing with respect to the Russian language were highlighted.