Abstract:
The manual creation of high-quality Math Word Problems presents a significant bottleneck for educators in low resource languages such as Sinhala and Tamil. While prior works have demonstrated the potential of multilingual models for this task, a systematic evaluation of advanced fine-tuning strategies to optimize performance remained a notable research gap.
This thesis presents a comprehensive experimental framework to address this gap. A series of nine controlled experiments were conducted on multilingual sequence-to-sequence models, primarily mBART, to thoroughly evaluate and compare various training methodologies. A mixed-method evaluation was employed, combining automated metrics (BLEU, METEOR) with a structured qualitative human evaluation.
The results prove that fine-tuning is a prerequisite for this task, as all models fail without it. The experiments further show that while a simple fine-tuning approach works, it fails to generalize across new problem domains and new languages. In contrast, advanced strategies like sequential transfer learning and multitask learning were validated as top-tier solutions that successfully overcome these critical limitations, producing high-quality and robust models. Sequential crosslingual transfer achieved the state-of-the-art performance. Finally, this research provides a clear, validated guide for optimizing MWP generation and offers a powerful new pathway for developing educational tools in underserved language communities.
