Predicting Machine Translation Performance on Low-Resource Languages: The Role of Domain Similarity (2402.02633v1)

Abstract: Fine-tuning and testing a multilingual LLM is expensive and challenging for low-resource languages (LRLs). While previous studies have predicted the performance of NLP tasks using machine learning methods, they primarily focus on high-resource languages, overlooking LRLs and shifts across domains. Focusing on LRLs, we investigate three factors: the size of the fine-tuning corpus, the domain similarity between fine-tuning and testing corpora, and the language similarity between source and target languages. We employ classical regression models to assess how these factors impact the model's performance. Our results indicate that domain similarity has the most critical impact on predicting the performance of Machine Translation models.