Modeling Emotional Expressions in Social Media Language Using BERT

This master's thesis was completed at the HSE University (National Research University Higher School of Economics), Faculty of Computer Science, in Moscow, 2026. The title of the dissertation is "Modeling Emotional Expressions in Social Media Language Using BERT". The total length of the work is 73 pages, excluding code and appendices. The bibliography includes 28 sources, and the text contains 23 figures and 5 tables. The dissertation consists of ten chapters: introduction, literature review, data description, data preprocessing, methodology, experiments, results and evaluation, discussion, conclusion, and references. The keywords are: emotion detection, BERT, social media, Twitter, volleyball, transformer architectures, multi-class classification, class imbalance, macro-averaged F1-score, and natural language processing. The object of the study is emotional expressions in short Twitter texts related to sports (specifically volleyball), while the subject is emotion classification methods based on traditional logistic regression and the transformer-based BERT model, including their comparative effectiveness and interpretability. The goal of the thesis is to develop and empirically compare the effectiveness of BERT and a classical machine learning method (logistic regression) for emotion detection in informal, noisy social media texts using volleyball-related tweets as a case study. To achieve this goal, the following tasks were carried out: collecting and manually annotating a dataset into seven emotion categories, fine tuning a BERT model, implementing a logistic regression baseline with TF IDF features, evaluating both models considering strong class imbalance (macro-averaged F1-score), and performing error analysis and model interpretation using explainable AI techniques. The relevance of the topic is driven by the growth of social media, the difficulty of automatic emotion analysis in short informal messages, and the lack of research in sports domains, especially volleyball. The scientific contribution of the work lies in creating an annotated dataset of volleyball tweets, conducting a comparative analysis of two approaches, and demonstrating BERT's advantage on imbalanced data. The practical value consists of a reproducible pipeline (collection → preprocessing → annotation → training → evaluation) that can be used for building emotion aware systems in sports analytics and social media monitoring. As a result of the study, BERT significantly outperformed logistic regression in macro-averaged F1-score; a strong class imbalance was identified (neutral and happiness together account for about 70% of the data); and a functional pipeline for domain specific emotion analysis was developed. The proposed approach can be integrated into fan sentiment monitoring systems, feedback analysis tools for sports organizations, and social listening platforms. Future research directions include exploring more powerful models (RoBERTa, T5), multimodal analysis (emojis, images) and extension to other sports and languages.