Urdu-LC: Length-Controlled Abstractive Summarization for a Low-Resource Language

Abstract

The rapid expansion of digital text has intensified the need for effective summarization techniques, particularly for languages with limited computational resources. Although recent progress in abstractive summarization has produced strong results for high-resource languages, comparable developments for Urdu remain constrained due to data sparsity, morphological variation, and syntactic ambiguity. Another persistent issue in existing systems is the absence of explicit length control, despite its importance in applications that require summaries of predetermined size, such as headlines or mobile notifications. In this study, we propose a length-controlled abstractive summarization framework tailored for Urdu. The approach builds on multilingual transformer architectures, specifically mT5 and mBART, and extends them through the integration of length tokens, EOS-aware optimization, and constrained decoding strategies. These components collectively enable the generation of summaries that align more closely with user-specified length requirements. Experimental evaluation on the Urdu portion of the XL-Sum corpus shows that the proposed models achieve competitive ROUGE and BERTScore results while exhibiting a substantial improvement in length adherence. Comparative analysis against existing baselines and state-of-the-art systems further demonstrates that the proposed framework offers practical advantages for summarization in low-resource settings.