ChatGPT's Translation Skills Parallel Most Human Translators (spectrum.ieee.org)

A recent study published in the IEEE Transactions on Big Data highlights the translation capabilities of large language models (LLMs), specifically comparing them to human translators. Researchers found that while only certified professionals with over a decade of experience consistently outperformed LLMs, models like GPT-4 exhibited translation quality comparable to junior and medium-level human translators. This marks a significant advancement for AI in the field of language translation, indicating that LLMs have reached near-human levels of accuracy for specific tasks. The study involved standardizing comparisons, where various translator experience levels were defined, and both humans and LLMs were evaluated on their translations of common and less common language pairings. GPT-4, for instance, averaged 3.71 major errors compared to 3.27 for junior and 3.30 for medium translators. Notably, while LLMs struggled with literal interpretations, human translators sometimes over-interpreted nuanced texts. This research suggests that as AI continues to improve, particularly with models capable of deeper reasoning, there may be a future where LLMs could rival senior human translators in more complex translation tasks.