Genre classification: A machine learning based comparative study of classical bengali literature

Abstract

Bengali literature, specifically classical Bengali literature has been a source of inspiration, a spark for paradigm-shifting revolutions, and the sole sustaining source of cultural thirst for hundreds of millions of people over many generations. Unfortunately, very few attempts have been made to analyze this never-ending collection of literary works from the luminary figures of Bengali literature. The availability of high-quality research-ready datasets comprising all the authenticated literary works has been a key obstacle in conducting NLP research, utilizing the most recent advancements in deep learning and large language models. Identifying the genre of a given text snippet is a key step in analyzing a vast collection of works comprising different styles, themes, and motivations from classical authors. From classifying previously unexplored archival documents to identifying and suggesting similar literary works for modern recommender engines, genre classification opens the door for many downstream and specialized use cases. In this project, we initiate an ambitious goal of compiling a comprehensive dataset of literary works from classical authors and eventually extending the collection to contemporary writers as well. We explore both classical methods such as Naive Bayes as well as LSTM and recent transformer-based models to classify genre from short text snippets. We concluded that fine-tuning pre-trained BERT models produced much higher accuracy than both classical and LSTM models.