Optimizing Resource-Scarce Language Question Generation on Bengali : A Comparative Study of Transformer-Based Models

The process of creating potential questions from text or other types of data, like documents, images, graphs, etc., is commonly referred to as Automatic Question Generation (AGQ) or simply Question Generation System. The objective of this research is to explore the application of Question Generation (QG) within the frequently asked questions (FAQ) domain, ultimately leading to the design and implementation of an automatic FAQ generation system. Question Generation studies in Finnish, Swedish, Turkish, and English face limitations due to poor machine-translated datasets, inadequate evaluation metrics, low text quality, limited results, and restricted tabular data use during COVID-19. Eliminating these limitations, we have considered this and employed a series of refined text-to-text encoder-only and encoder-decoder transformer-based models named IndicBart, SahajBart, Biswabangla, BanglaBERT, BanglaT5, and lastly Google’s mT5-base model. The proposed bn-TOK + mT5-base + Beam Search model scores Rouge 37.225 and SacreBLEU 14.202 with METEOR 0.216 and demonstrates superior performance in given metrics, balancing lexical accuracy, structural coherence, and semantic meaningfulness when the context of the text and the answers are given to create questions that seem humane.