Hybrid rank-MMR

Text summarization is a challenging problem in natural language processing (NLP), especially for low-resource languages like Vietnamese or specialized documents (such as scientific articles). Currently, extractive text summarization models often rely on semantic matching using pre-trained language models or select sentences based on their position. However, these methods prove ineffective when applied to different languages or domains in extractive text summarization tasks. Therefore, we propose Hybrid Rank-MMR, a re-ranking model that combines semantic similarity with two features: semantic density and sentence position. This method can handle both news (relying on sentence position) and complex datasets (requiring semantic understanding). Furthermore, we refine the MMR algorithm to automatically control diversity and remove duplicate sentences during summary generation. We test on seven datasets in both English (CNN-DailyMail, Multi-News, XScience, WCEP) and Vietnamese (VNDS, ViMs, AbMuSu). The results show that this model significantly improves performance compared to the basic models. ROUGE-1 increases by +2.56 on Multi-News and +6.68 on ViMs. Finally, we integrate an explainable artificial intelligence (XAI) module based on large language models (LLMs). This module uses natural language to explain to the user why the system selected those sentences for summarization. Compared with prior zero-shot extractive systems such as MatchSum, which are applied without fine-tuning, our framework requires no parameter updates on any benchmark.