Testing dataset distribution.

The rapid proliferation of online news demands robust automated classification systems to enhance information organization and personalized recommendation. Although traditional methods like TF-IDF with Naive Bayes provide foundational solutions, their limitations in capturing semantic nuances and handling real-time demands hinder practical applications. This study proposes a hybrid news classification framework that integrates classical machine learning with modern advances in NLP to address these challenges. Our methodology introduces three key innovations: (1) Domain-Specific Feature Engineering, combining tailored n-grams and entity-aware TF-IDF weighting to amplify discriminative terms; (2) BERT-Guided Feature Selection, leveraging distilled BERT to identify contextually important words and resolve rare-term ambiguities; and (3) Computationally Efficient Deployment, achieving 95.2% of the accuracy of BERT at 1/52.4th of the inference cost. Evaluated on a balanced corpus of Sina News articles in 11 categories, the system demonstrates a test precision of 95.12% (vs. 84.43% for SVM+TF-IDF baseline), with statistically significant improvements confirmed by 5-fold cross-validation(p