From Extractive to Transformer-Based Summarization: A Review of Low-Resource Language Approaches

The digital world is expanding at an unprecedented pace, with recent estimates suggesting that over 463 exabytes of data are generated daily, nearly 80-90% of which remains unstructured. In this rapidly evolving landscape of textual information, the ability to distill large volumes of content into concise, meaningful, and context-rich summaries has become increasingly critical. Over time, text summarization has progressed beyond traditional extractive techniques, propelled by developments in large language models (LLMs), pre-trained language models (PLMs), and deep neural networks, enabling more coherent, fluent, and context-aware outputs. Despite these advances, performance remains inconsistent across languages and domains, particularly in settings where annotated resources are scarce and linguistic diversity is high. With more than 7,000 languages spoken worldwide and only a small fraction adequately supported in NLP, low-resource summarization continues to gain attention as a critical research area.