Title: Bridging the Gap: Effective Prompting Techniques for African Natural Language Inference

Abstract:

While Large Language Models (LLMs) are increasingly tested in multilingual contexts, their reasoning capabilities in low-resource African languages remain largely unexamined, particularly when relying solely on prompt-based inference without fine-tuning. This paper presents a comprehensive analysis of various prompting strategies for Natural Language Inference (NLI) applied to Swahili, Yoruba, and Hausa, utilizing the AfriXNLI benchmark. Our evaluation covers five distinct prompting methods—Baseline (zero-shot), Script-Aware, Language Specific, Contrastive, and Native-Label Self-Translation (NL-STP)—across two mid-sized open-weight models: Llama3.2-3B and Gemma3-4B.

To strictly isolate the impact of prompt design, this study excludes the influence of few-shot examples and Chain-of-Thought reasoning. The results reveal substantial variations in performance across different classes depending on the strategy employed, with certain configurations exhibiting significant prediction skew and a tendency toward class collapse, particularly for neutral labels. Among the tested approaches, Contrastive prompting emerged as the most consistent and robust method, demonstrating steady improvements across both languages and models while offering a superior balance between class behavior and overall accuracy gains.

Notably, the study indicates that carefully designed prompts can outperform more powerful baseline models that incorporate few-shot and Chain-of-Thought techniques. These findings underscore the critical importance of prompt formulation for multilingual NLI in low-resource settings, suggesting that language-aware decision structuring can significantly enhance model robustness in resource-constrained environments.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC