Advanced News Classification Model with Capsule Networks Through Word2Vec and BERT Embeddings

This paper explores the application of Capsule Networks (CapsNet) for news classification, addressing the limitations of traditional deep learning models such as RNN, LSTM, an GRU. The BBC News dataset from Kaggle was utilized, where text preprocessing involved tokenization, stopword removal, and lemmatization. To enhance feature representation, Word2Vec and BERT embeddings were applied to convert raw text into meaningful vector representations. Unlike conventional models that rely on max pooling, CapsNet employs a dynamic routing mechanism, preserving spatial hierarchies and word dependencies. The classification performance of CapsNet was evaluated against baseline models (RNN, LSTM, and GRU) using accuracy, precision, recall, and F1-score. Experimental results indicate that CapsNet consistently outperforms sequential models, particularly when combined with BERT embeddings, achieving an accuracy of 96.2%, surpassing GRU’s 93.4%. The findings demonstrate the effectiveness of CapsNet in retaining contextual relationships, reducing misclassification, and enhancing text classification accuracy. These results highlight Capsule Networks as a promising alternative for automated news classification, contributing to advancements in natural language processing and text categorization.