TWO-STAGE ALGORITHM FOR AUTOMATED OPINION EXTRACTION FROM RUSSIAN POLITICAL TEXTS

The article discusses neural network approaches for automated opinion extraction in Russian. A two-stage algorithm is proposed based on solving particular problems of opinion target extraction (subject extraction) and target sentiment analysis using machine learning models of the transformer architecture The datasets of Russian political texts from microblogs, news publications and public speeches are presented, based on the existing Russian-language sets CABSAR and RuSentNE, as well as the English-language set NewsMTSC, translated using machine translation. The OTE_Ru_dataset includes sentence tokens markup by opinion target in BIO format. The TSA_Ru_dataset includes targeted emotional markup in relation to the subjects of opinions in the context of sentences in the author's format. Additional training of models of the DeBERTa-base, ruRoPEBert-e5-base-512, ruBert-large, ruElectra-large, ruRoBERTa-large, XLM-V-base architectures, previously trained in texts in Russian, was carried out. Metrics for evaluating the quality of machine learning models are described. High performance has been achieved by models whose tokenizers take into account the relative position of the word in the text. The best results of target extraction were achieved by the DeBERTa-based model. The best target sentiment analysis classifier is ruRoBERTa-based classifier. The applicability of the research results to applied tasks such as the analysis of socio-political processes and the identification of ideologemes is described.