Reference Translation

Reference translation, the process of creating high-quality human-translated texts used to evaluate machine translation systems, is undergoing significant re-evaluation. Current research focuses on improving the reliability and quality of reference translations themselves, exploring methods to create "optimal" references and using techniques like multiple reference translations and synthetic data augmentation to enhance training data. This work leverages large language models (LLMs) like GPT-4 for both evaluating translation quality and generating improved references, aiming to address limitations of existing evaluation metrics and ultimately improve the accuracy and robustness of machine translation systems. The resulting advancements have implications for both the development of more sophisticated machine translation models and the broader field of computational linguistics.