Researchers from Tencent AI and the Harbin Institute of Technology introduced TasTe, a method for teaching large language models (LLMs) to translate through self-reflection. LLMs have shown strong performance in natural language processing tasks, including machine translation, but still lag behind supervised neural machine translation systems in quality. TasTe aims to improve LLM translation capabilities by incorporating a self-reflection process.
The TasTe framework involves two stages: in the first stage, LLMs generate preliminary translations and self-assess the quality of these drafts. The models then refine their translations based on this evaluation to produce final translations. Low-quality drafts undergo extensive modifications, while high-quality drafts require minimal changes. This process mirrors the human “try-evaluate-improve” approach to complex tasks.
TasTe was evaluated in four language directions using the WMT22 benchmark, outperforming existing methods by enhancing translation quality through self-assessment. The approach was also tested as an automatic post-editing (APE) tool, showing effectiveness in refining translations generated by other systems.
The researchers provide their code and datasets for further research on GitHub. The authors of the paper are Yutong Wang, Jiali Zeng, Xuebo Liu, Fandong Meng, Jie Zhou, and Min Zhang. TasTe not only improves LLM translation quality but also serves as an effective APE tool for enhancing translations from other systems.
Source link
Source link: https://slator.com/how-to-teach-large-language-models-to-translate-through-self-reflection/