On Multilingual Text Summarization: A Review of Model Architectures, Techniques, and Research Gaps

Abstract

The rapid expansion of multilingual digital content has intensified demand for automated text summarization systems that can operate across diverse languages. Yet the architectural and methodological foundations of this domain remain dispersed across individual studies. This paper presents a systematic literature review of multilingual text summarization research in Natural Language Processing, with particular attention to model architectures, multilingual enabling techniques, and existing research gaps. Following a PRISMA-guided review protocol, 47 studies published between 2017 and 2026 were identified, screened, and synthesized from major scholarly databases using predefined eligibility and quality assessment criteria. The analysis shows that the literature is dominated by four dominant architectural paradigms: multilingual sequence-to-sequence models, large language models, adapter-based models, and hybrid architectures, with a recent shift toward large language model-based approaches. Across these studies, multilingual capability is primarily achieved through pre-training and transfer learning, with support from parameter-efficient adaptation and data augmentation. The review further identified four persistent limitation categories: low-resource language underrepresentation; dataset and benchmark constraints; evaluation limitations dominated by ROUGE overreliance; and generalization and scalability challenges, manifested as the curse of multilinguality. These findings indicate that current progress in the field is largely driven by scaling existing models rather than resolving structural multilingual limitations, particularly those related to data availability and model architecture. This paper provides a consolidated view of architectural and methodological trends and offers a structured, informed basis for future research on equitable multilingual summarization, particularly for low-resource languages.