Prompt Engineering vs. Fine-tuning for Arabic Word Sense Disambiguation: An Empirical Study with Generative LLMs

Word Sense Disambiguation (WSD) is a foundational yet challenging Natural Language Processing (NLP) task that aims to determine the correct meaning of an ambiguous word given its context, with direct applications in machine translation, information retrieval, and question answering system. Arabic presents exceptional disambiguation difficulty due to its rich morphology, extensive polysemy, and the absence of diacritical marks in written text, leaving the correct interpretation of a word entirely dependent on context. Over the years, Arabic WSD has attracted growing research attention, progressing from knowledge-based and supervised machine learning to the current dominance of transformer-based models trained on Arabic corpora. Although fine-tuned transformer models have shown competitive results on Arabic WSD benchmarks, they provide a major practical challenge since they require large amounts of labeled training data and computing resources, which remain inaccessible in many Arabic NLP settings. This work proposes an empirical comparison of prompt engineering strategies against finetuned models for Arabic WSD using generative Large Language Models (LLMs). It evaluates multiple techniques of increasing sophistication such as zero-shot, few-shot, chain of thought and definition augmented prompting across different generative LLMs including GPT-4o, Qwen, LLama, and ALLaM and comparing their performance against their fine-tuned version on established WSD datasets. The findings are intended to provide Arabic NLP practitioners with a principled empirical basis for navigating the performance-cost tradeoff in Arabic WSD system design, advancing the goal of making high-quality Arabic language understanding more accessible in resource-constrained environments.