Prompt Engineering for Low-Resource NLP: A Survey with Urdu-English Applications

Abstract

With limited computational resources and labeled data, large language models (LLMs) can be effectively adapted to new tasks through prompt engineering especially in low-resource language environments. Focusing on natural language processing (NLP) in Urdu which is a language with complicated morphology, script variations, code-mixing and limited high-quality datasets, this study investigates prompt engineering strategies. Prompting techniques such as zero-shot and few-shot are methodologically structured and examined especially for Urdu-specific issues including tokenization, linguistic grounding and cross-lingual evaluation. In addition to the survey, a prompt-only experimental investigation is carried out on Urdu abstractive summarization and sentiment classification. Multilingual (mGPT, mBART), generalist instruction-tuned (GPT-4, Gemini, Gemma) models and an Urdu-specific (urT5) model are evaluated using a single prompt-based architecture. For cross-lingual summarization, a two-stage pipeline is used involving urT5 generates summaries in Urdu that are then translated into English using mBART for evaluation. Clear task-dependent performance differences between model variants are revealed by quantitative results obtained using common automatic measures. The findings, with Urdu-specific models demonstrating stronger alignment in generation-centric tasks and generalist models remaining more effective for instruction-driven classification, show that prompt engineering is a valuable but non-uniform adaptation method for Urdu NLP, along with providing practical guidance for future low-resource language research