HSF-Urdu: A Hybrid Semantic Fusion Framework for Multi-Document Urdu News Summarization Using Large Language Models

Abstract

The increasing volume of online news published by multiple media organizations has created a growing need for effective multi-document text summarization systems capable of generating concise and coherent summaries. While Large Language Models (LLMs) have shown significant success in abstractive summarization, their direct application to multi-document summarization can lead to a few limitations like repetition of content, partial event coverage, and poor utilization of context, especially in low-resource languages like Urdu. In this paper, we propose a Multi-Document Urdu News Summarization using Hybrid Semantic Fusion Framework (HSF-Urdu), which uses semantic sentence ranking, context fusion, and prompt-guided LLM for summarization. The framework first performs pre-processing and clustering of news articles by events, then uses semantic representation and ranking to identify informative and non-redundant sentences, and then to build an integrated contextual representation for generating a summary using an LLM. The proposed approach was tested on Urdu multi document news data set gathered from various Pakistani newspapers using ROUGE, BLEU and BERTScore with qualitative analysis and comparison with representative baseline methods. Experimental results show that HSF-Urdu produces coherent and semantically consistent summaries, which also provide significantly more information coverage and less redundancy than the current approaches. The results emphasize the potential of semantic information processing combined with LLM-based Urdu summarisation for low resource languages.