Method for Detection of Disinformation Based on Text Data Analysis Using TF-IDF and Contextual Vector Representations

The article considers an approach to detecting fake news in the digital environment through text analysis using machine learning and natural language processing methods. The proposed method is based on a hybrid text representation combining frequency features (TF-IDF) and contextual embeddings obtained using the IBM Granite model. A complete data processing cycle was developed, covering the stages of exploratory analysis (EDA), text preprocessing and tokenization, forming vector representations, training a logistic regression model, and obtaining key metrics. The main stages of text preprocessing  included  converting  all  characters  to  lowercase,  removing  URLs  and  HTML  tags, cleaning from non-letter characters and excess spaces, eliminating duplicates to avoid re-training, and unifying the values of specific fields. A combination of TF-IDF with contextual embeddings was used to vectorize  the  cleaned  texts,  which  allowed  the  model  to  simultaneously  consider  the  statistical significance of terms and their semantic context within the messages. The constructed logistic regression model combined with a hybrid representation of text data demonstrated high efficiency, achieving an overall accuracy of 82 % and balanced F1-measure values for the “true” and “fake” classes. An analysis of TF-IDF feature weights based on logistic regression coefficients was applied to identify the most relevant  terms.  The  study  showed  that  the  model  tends  to  associate  truthful  information  with Ukrainian-language, neutral vocabulary, while texts with signs of disinformation often contain Russian- language elements characteristic of propaganda or manipulative messages. Further research will be