A NOVEL MT5-BASED FRAMEWORK FOR ABSTRACTIVE TEXT SUMMARIZATION IN THE LOW-RESOURCE SINDHI LANGUAGE

Abstract

The rapid growth of digital textual information has increased the demand for automated summarization systems capable of generating concise, coherent, and semantically meaningful summaries. Although transformer-based abstractive summarization has achieved considerable success in high-resource languages, its application to low-resource languages remains constrained by limited annotated datasets and inadequate linguistic representation in multilingual pretrained models. Sindhi, despite its widespread use across South Asia, remains underexplored in abstractive summarization research. To address this gap, this study proposes an mT5-based abstractive text summarization framework specifically designed for the Sindhi language. A dedicated corpus comprising 6,500 article-summary pairs was constructed from multiple digital news sources and processed using language-specific normalization and tokenization techniques adapted to the Perso-Arabic Sindhi script. The proposed framework was fine-tuned and evaluated against established extractive summarization baselines, including LexRank and TextRank, using ROUGE, BERTScore, and qualitative human evaluation. Experimental results demonstrate that the proposed framework achieved ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.51, 0.26, and 0.47, respectively, with a BERTScore F1 of 0.8975. The framework outperformed the evaluated extractive approaches in lexical and semantic evaluation. Human assessment further indicated its ability to generate fluent, coherent, and contextually relevant summaries through semantic abstraction and paraphrasing. These findings establish a foundation for Sindhi abstractive summarization and demonstrate the potential of multilingual transformer adaptation for low-resource natural language processing.