This paper presents a comprehensive analysis of text classification techniques for automating news categorization in a Romanian television station. We constructed an original dataset of approximately 24,000 labeled news stories collected from the station’s media asset management system over a 15-month period. Our research evaluates the performance of seven supervised machine learning classification algorithms: Multinomial Naive Bayes, Support Vector Machine, Logistic Regression, Random Forest, XGBoost, k-Nearest Neighbors, and Multi-Layer Perceptron, combined with three feature extraction methods: TF-IDF, CountVectorizer, and HashingVectorizer. We tested these combinations on both the original imbalanced dataset and a balanced version created using random oversampling. Results demonstrate that balancing the dataset significantly improves classification performance across all models. On the balanced dataset, Random Forest and Multi-Layer Perceptron achieve near-perfect performance regardless of the feature extraction technique employed. The study reveals that while feature extraction choice notably impacts performance on imbalanced data, these differences become minimal after balancing. Our findings offer valuable insights for building automated news categorization systems for Romanian content, demonstrating that established NLP techniques can be effectively applied to this less-resourced language with appropriate preprocessing and balancing methods. This approach could allow news organizations to save time and resources while improving categorization consistency.