Bangla Book Genre Classification: An Advanced Multi-Class Approach Using Transformer-Based Models

This paper explores a diverse range of multiclass classification methods for Bangla books, particularly those accessible on digital platforms. Despite being one of the most widely spoken languages with a rich literary heritage, Bangla remains underexplored in content classification due to the lack of a comprehensive dataset. We introduce an expanded dataset of 46,429 books classified into 14 distinct categories. The categories are Education, Fiction, Non-Fiction, Poetry, Children’s Book, History, Biography, Literature, Adventure, Crime, Sci-Fi, Programming, Others. In order to achieve this, we used advanced machine learning models like BanglaBERT, XLM-RoBERTa and a CNN-LSTM attention-based deep learning approach. Our models demonstrate substantial improvements in classification accuracy, attributed to the increased dataset size and the adoption of refined model architectures. Both attention base and transformer based have outperformed previous benchmarks, with accuracy leaping from approximately around 16% to around 75% a direct result of the novel dataset we developed.Among the models that we have tested, BanglaBERT stands out in terms of stability, showcasing exceptional performance due to its advanced transformer-based architecture. This positions BanglaBERT as the most effective model for Bangla content classification in our study, highlighting its superiority in handling the linguistic complexities of Bangla text.