EVALUATING LARGE LANGUAGE MODELS (LLMS) FOR TRANSLATING ENGLISH NOUN PHRASES TO YORUBA

This study examines the ability of large language models (LLMs) to translate English noun phrases into Yorùbá, a low-resource African language characterised by complex tone and morphology. While LLMs like ChatGPT and Claude have demonstrated general success in multilingual tasks, their ability to handle specific linguistic features, such as number, possession, and tonal marking in Yorùbá, has not been well-studied. This research addresses that gap by focusing specifically on the translation of noun phrases. The study adopted both quantitative and qualitative methods. One hundred English sentences and their Yorùbá equivalents were collected from JW.ORG. From these, 50 noun phrases were extracted and translated using eleven LLMs: ChatGPT, Claude, Google Gemini, Microsoft Copilot, AI Chat, POE AI, Perplexity, YouChat, SIDER AI, Monica, and MUME AI. The translations were then compared with the human-translated Yorùbá phrases from the original JW.ORG corpus. The analysis followed a typological and error-based approach, identifying patterns of mistranslation and structural deviation. The study also applied elements of translation theory and LLM prompting techniques. The results show that while LLMs are capable of producing fluent Yorùbá, many translations fail to capture key grammatical features such as plurality, definiteness, and tone. Some models, such as Claude and ChatGPT, performed better than others, particularly in terms of lexical accuracy and sentence fluency. However, even the best-performing systems frequently omitted tonal markings and misunderstood possessive or complex noun phrases. Translation errors were more common in culturally specific or semantically dense phrases. In conclusion, LLMs still have significant limitations in handling morphosyntactic detail and tonal clarity in Yorùbá noun phrase translation. The study recommends the creation of diacritic-sensitive evaluation tools, fine-tuning LLMs on Yorùbá corpora, and integrating human post-editing or hybrid systems. These steps would improve translation accuracy and promote the development of language technology for under-resourced African languages. Keywords: Large Language Models (LLMs), Yoruba Translation, Machine Translation, Morphosyntactic Analysis, Low-Resource Languages