SENTIMENT ANALYSIS OF TEXTS USING RECURRENT NEURAL NETWORKS OF THE TRANSFORMER ARCHITECTURE

This study focuses on the automated process of determining sentiment (emotional coloring) in text messages from Telegram channels to enhance Ukraine’s information security. The principal challenge addressed lies in the need for rapid and accurate detection of negative, positive, or neutral messages within large-scale data streams without additional fine-tuning on local datasets. The essence of the results obtained is the implementation of a zero-shot classification approach, based on the multilingual transformer model XLM-RoBERTa, which in the experiment yielded the following metrics: Accuracy = 0.4718, Precision = 0.7138, Recall = 0.4718, and F1 Score = 0.5044. Owing to the model’s strong ability to generalize lexico-semantic patterns, a stable compromise between Precision and Recall was achieved, thereby increasing the efficiency of message analysis in large data volumes. These results are explained by the architectural features of XLM-RoBERTa, primarily its multilingual nature and deep layer structure, which enable proper handling of multilingual texts without dedicated local training. Conclusions. The proposed approach is advisable when there is a large, diverse corpus of data requiring prompt detection of potential negative informational influences and timely counteraction. Practically, this significantly reduces the time spent on manual monitoring of the information space and eases the burden on analysts, thereby strengthening the ability of organizations or information security units to respond rapidly to destructive content. The research results can also be integrated into decision support systems, serving as a foundation for the development of software solutions aimed at monitoring the information space.