INFORMATION TECHNOLOGY FOR UKRAINIAN-LANGUAGE FAKE NEWS DETECTION IN SOCIAL NETWORKS CYBERSPACE BASED ON MACHINE LEARNING AND NATURAL LANGUAGE PROCESSING METHODS

Abstract

The article addresses the problem of automated fake news detection in the Ukrainian-language information space, which has become particularly relevant under conditions of hybrid warfare and the intensive use of disinformation as a tool of information influence. The aim of the study is to develop and experimentally evaluate an effective system for classifying Ukrainian-language news texts using natural language processing and machine learning methods. The introductory section substantiates the relevance of the research topic, defines the object and subject of the study, and formulates its goal and main objectives. The related work section provides an overview of existing approaches to fake news detection, including classical machine learning algorithms, deep neural networks, and transformer-based models, and highlights their limitations in the context of the Ukrainian language.

The theoretical section systematizes methods of automated text analysis and identifies linguistic features of Ukrainian news content, as well as challenges related to the lack of large annotated corpora and language-specific resources. The methodology section describes the complete research pipeline, including data collection, text preprocessing, cleaning, tokenization, and lemmatization using tools adapted for Ukrainian, as well as the construction of semantic vector representations based on the Sentence-BERT model. Several machine learning classifiers, namely Logistic Regression, Random Forest, Support Vector Machine, and XGBoost, were implemented and compared using cross-validation and hyperparameter optimization.

The results section presents an experimental evaluation of the models using accuracy, precision, recall, F1-score, and ROC-AUC metrics. The findings demonstrate that the Support Vector Machine model achieves the best performance, reaching a classification accuracy of 93.2% even under class imbalance conditions. Computational efficiency and runtime characteristics of the proposed approach are also analysed, along with its potential for practical deployment. The conclusions summarize the main outcomes of the study, confirm that the research objectives have been achieved, and outline directions for future work, including adaptation to a multilingual environment and improving model interpretability. The proposed system can be applied in media monitoring, fact-checking initiatives, and tools aimed at strengthening information security.