Contextual deep learning for accurate news article categorisation with pre-trained embeddings

Abstract An increasing amount of online news content in digital journalism leads to novel and complicated issues regarding its classification and organisation. Systems that automate operational tasks can provide numerous advantages over systems that manually classify documents. The current work attempts to solve this problem using contextual and semantic deep learning. The text’s semantic meaning is captured using pre-trained word embeddings, and understanding is aided by a hybrid neural structure that incorporates local and distant text dependencies. This technique is compared against classical machine learning on two prominent news datasets. The deep learning approach yields a remarkable improvement in classification, reaching over 91% accuracy on the AG News corpus, a balanced four-class English benchmark, while its performance on the News Category Dataset V3, which is more complex and highly unbalanced, is considerably lower. These results highlight the effectiveness of contextual and semantic modelling for news categorisation, while also illustrating the impact of dataset complexity and class imbalance on achievable performance.