Automatic Irony and Sarcasm Detection in Russian Sentences: Baseline Methods

The paper describes experiments performed on two sets of manually annotated data. The task of irony and sarcasm detection in Russian sentences was solved using baseline classifiers, i. e., BERT, Bi-LSTM, SVM, Random Forest, Logistic Regression. The best achieved F1-score for each classifier was 0.76, 0.73, 0.66, 0.64, 0.68 respectively. The results achieved by BERT and Bi-LSTM classifiers are comparable with the results from the articles describing the application of similar approaches for English language. Analysis of the results allowed to conclude that transferring the word context improves classification metrics and refinement of training data allows to improve the classifier's performance.