When Graphs Meet Fine-Tuning: A Dual Approach to Low-Resource Assamese QA

This paper introduces a novel framework designed to tackle question-answering tasks, with a focus on low-resource languages: Assamese. The framework employs two distinct methodologies: an unsupervised graph-based approach that leverages Named Entity Recognition, graph embeddings, and similarity techniques, alongside a supervised approach that finetunes pretrained models to enhance performance. We created a dataset which consists of $\mathbf{1 0 0}$ distinct contexts, each containing $\mathbf{1 0}$ extractive questions and their corresponding answers, all placed sequentially within the context. Through rigorous experimentation, we demonstrate the effectiveness of these methodologies. The Muril model, a leading pretrained model for Indic languages, achieved an exact match score of 62 and an F1 score of 75.44, making it the top performer. On the other hand, the unsupervised graph-based approach yielded a promising F1 score of 45.92, showcasing its potential for domain-specific tasks in low-resource settings. These findings contribute to the field of Natural Language Processing by offering scalable solutions for developing question-answering systems, particularly for underresourced languages.