Evaluating Multilingual Models: A Comparative Evaluation of MT0 and MT5 for Urdu Text Articles Summarisation

Abstract

Abstractive text summarization is a critical task in natural language processing aimed at generating concise summaries while preserving the core meaning of the original text. This paper focuses on abstractive summarization of Urdu articles using the MT5 pre-trained model, a variant of T5 that supports multilingual capabilities. Specifically fine-tuned for Urdu, MT5 harnesses cross-linguistic knowledge to enhance summarization quality. We evaluate its performance against the MT0 model, a baseline, using ROUGE metrics. Our experiments, conducted on a novel Urdu dataset, demonstrate significant improvements in summary quality with MT5 compared to MT0. This research underscores the utility of multilingual models in advancing abstractive summarization and highlights avenues for future development in natural language processing for underrepresented languages.