Advancing Bangla NLP: Transformer-Based Question Generation using Fine-Tuned LLM

This research explores the development of a question generation model for Bangla text using the Bangla T5 base model, a transformer-based architecture tailored for the language. The study addresses the underrepresentation of Bangla in natural language processing (NLP) by leveraging a carefully curated dataset of Bangla paragraphs and corresponding questions. The Bangla T5 model, pre-trained on a large Bangla corpus, is fine-tuned to generate contextually relevant and coherent questions from input paragraphs. The model’s performance is evaluated using standard metrics such as Exact Match (EM) and ROUGE, achieving a 28.65% Exact Match score. The ROUGE scores demonstrate strong alignment with the reference questions, with ROUGE-1 F1 at 73.64%, ROUGE-2 F1 at 57.86%, and ROUGE-L F1 at 70.77%. These results highlight the model’s ability to capture the linguistic nuances of Bangla and generate questions that are semantically and syntactically accurate. This work contributes to the advancement of Bangla NLP and underscores the potential for extending similar methodologies to other low-resource languages, promoting inclusivity and linguistic diversity in AI systems.