Cross-Lingual NLP and Large Language Models for Low-resource Languages: A Review

Natural Language Processing (NLP) is being revolutionized by the increasing competence of large language models (LLMs) and sophisticated multilingual transfer methods. While there has been impressive progress in machine translation, cross-lingual comprehension, and zero-shot/few-shot learning, low-resource languages (LRLs) continue to be technologically underdeveloped. This overview integrates and examines critically findings from forty newly published peer-reviewed studies within five broad thematic categories: cross-lingual transfer, unsupervised/semi-supervised machine translation (MT), adapter and prompt-based models, data augmentation techniques, and resource building and benchmarking. we draw methodological and conceptual perspectives from CSS scholarship on LLMs to investigate synergies in solving LRL problems. The review considers technical architectures like machineagnostic data-gated(G)/machine-agnostic data-extended(X) adapters, multilingual prompt translation (MPT), and hybrid neural-symbolic models, as well as the sociotechnical consequences of using LLMs in resource-poor settings. The review shows a proper search and selection method inspired by PRISMA, covering works that are categorized mainly in five thematic clusters: cross‑lingual transfer, unsupervised and semi‑supervised machine translation, adapter and prompt‑based architectures, data augmentation, and resource building and benchmarking. The studies are analyzed on data setting, target low-resource language and evaluation metrics helping in understanding where multilingual LLMs help the most.