Sentiment Analysis of Russian Reviews to Estimate the Usefulness of Drugs Using the Domain-Specific XLM-RoBERTa Model

This paper considers the problem of classifying Russian-language drug reviews into five classes (corresponding to the authors' rating scores) and into two classes (the drug is helpful or useless) in terms of the sentiment analysis task. The dataset of reviews with a markup of pharmaceutically-significant entities and the set of neural network models based on language models are used for the study. The result obtained in this task formulation is compared to a solution based on extracting the named entities relevant to the drug-taking effectiveness, including the positive dynamics after the drug, the side effects occurring, the worsening of the condition, and the absence of the effect. It is shown that both approaches (a classification one and one based on the extracted entities) demonstrate close results in the case of using the best-performing model - XML-RoBERTa-sag.