A hybrid model for automatic news message classification

This paper presents a hybrid model for automatic news classification that combines a pre-trained BERT encoder with a parallel convolutional stream of 2-, 3- and 4-gram filters operating over a shared trainable embedding. The model was evaluated on the AG News benchmark, six further public datasets and a purpose-built noisy corpus of 15,340 articles collected from BBC News, Reuters and Al Jazeera RSS feeds. On AG News the hybrid model reached 95.2% accuracy and a macro F1 of 0.951, 1.7 pp above the BERT-only baseline; on the noisy corpus, together with weighted cross-entropy loss, label smoothing and back-translation, the gain grew to 7.8 pp. The paper also reports a systematic ablation study, an error analysis of the Science/Technology–Business boundary, post-hoc temperature scaling and a boundary-aware re-ranker that further improves macro F1 by 0.7 pp.