Attention Based Bengali News Classification System Using Hybrid Deep Learning Model

The headline is often regarded as the soul of a news article, as it conveys the essence of the story and captures the attention of the reader. Many readers decide whether to engage with a news article based on the headline alone. This study presents a novel automated system for categorizing Bangla news articles into six predefined categories: International, Sports, National, IT, Amusement, and Politics. To address the growing need for effective information filtering in the Bangla online news landscape, this research employs a large-scale document repository comprising 136,811 news articles, each annotated with three features to fine-tune deep learning (DL) models for Bangla text. Specifically, a stacked convolutional layer and attention layer are integrated with BiLSTM to enhance performance. The proposed attention-layer-based CNN+BiLSTM model is capable of capturing sequential relations and contextual features in the text, which is essential for accurate classification. Experimental results demonstrate that the optimized BiLSTM model outperforms GRU models, achieving an accuracy of 84.04% compared to 83.91% for GRU. Moreover, the attention-layered BiLSTM model takes less training time than GRU and adjusts to the elaborate patterns in the dataset more successfully.