Classification of Emotional State Based on Text Data Using Deep Learning Methods

This thesis is devoted to the quantitative analysis of Bank of Russia communication using natural language processing methods. The study constructs and annotates a corpus of official Bank of Russia texts, including key rate press releases, statements, and summaries of key rate discussions. A multidimensional annotation framework is proposed, in which each sentence is classified by topic, temporal orientation, and monetary policy sentiment. The thesis compares classical machine learning models based on TF-IDF features with transformer-based models for Russian-language text classification. The results show that topic classification and temporal-orientation classification are largely driven by lexical features, whereas monetary policy sentiment classification requires a more context-sensitive approach. Based on the selected model, quantitative indicators of Bank of Russia communication are constructed, including the Net Policy Sentiment index, the Straightforwardness Index, the Explanation Index, and the Net Confidence Index. An exploratory econometric analysis is then conducted to examine the association between the constructed communication indices and macroeconomic variables, including the key rate, RUONIA, inflation, the exchange rate, and GDP dynamics. All data used in this thesis were collected from open-access sources, primarily from the official website of the Bank of Russia. The corpus does not contain personal data, and the analysis is conducted at the level of publicly available institutional documents. Therefore, the study does not raise privacy-related concerns.