In a September 10, 2024 paper, researchers from Google introduced a multi-step process aimed at improving the translation quality in large language models (LLMs) by mimicking human translation workflows. They explained that machine translation (MT) has been treated as a “single, monolithic task,” where a source text is simply translated into a target language. The Google researchers argue, however, that translation is a “multi-faceted process encompassing several sub-tasks that navigate a bilingual landscape.” They emphasized that “recent advancements in large language modeling offer promise for re-defining MT to align more closely with human translation processes.” To that end, they proposed a framework that engages LLMs in a multi-turn, step-by-step process consisting of four distinct stages: pre-translation research, drafting, refining, and proofreading. The process starts with the LLM being prompted to conduct background research to identify potential challenges in translating the source text. Next, the drafting phase focuses on creating an initial translation that prioritizes fidelity to the original text. This draft is then revised to enhance its fluency and coherence. Finally, the proofreading phase ensures that the translation is polished and free of errors. By integrating both pre-translation (research) and post-translation (refinement and proofreading) stages into a single framework, the Google researchers aim to “streamline the translation process” while relying solely on the LLM’s internal knowledge, eliminating the need for external resources. The proposed framework draws inspiration from the chain-of-thought prompting technique used in LLMs. By breaking down the translation task into smaller, manageable steps, the model can generate more accurate and contextually appropriate translations. The researchers tested their approach using the Gemini 1.5 Pro model on long-form documents (document-level translation) across ten languages, including Chinese, Ukrainian, Russian, Japanese, Hebrew, Czech, German, Hindi, Icelandic, and Spanish. They compared their method against traditional zero-shot translation techniques — where the model is instructed to translate the source text directly — and earlier human-like LLM-driven approaches using automatic evaluation metrics. They found that the translations generated through the step-by-step process outperformed traditional zero-shot translations, particularly in document-level translations where context is crucial. “Our approach improves translation quality over directly translating the entire document with a single prompt,” they said. The researchers highlighted the importance of both the pre-translation research and post-translation refinement stages, noting that the most substantial quality improvements occurred when these two stages were combined. “Those stages bring complimentary benefits,” they said. “Our findings highlight the potential of LLMs to progressively improve their translations, moving beyond the traditional view of machine translation as a monolithic sequence mapping task,” the researchers concluded. Authors: Eleftheria Briakou, Jiaming Luo, Colin Cherry, Markus Freitag