Reproduce “EnAnchored-X2X: English-Anchored Optimization for Many-to-Many Translation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sen Yang, Yu Bao, Yu Lu, Jiajun Chen, Shujian Huang, Shanbo Cheng
Researchers from Nanjing University and ByteDance Research developed EnAnchored-X2X, a framework that enhances large language models' direct non-English-to-non-English (x2x) translation by leveraging English-centric strengths. The method improved x2x BLEURT scores by up to 5.49 points for Llama2-7B and showed collateral improvements in English-centric translation.
Key insights
- LLMs' established English-centric translation capabilities can be strategically leveraged to effectively bootstrap and significantly improve their x2x translation performance.
- Optimizing LLMs for x2x translation through an English-anchored, preference-based approach can lead to unexpected collateral improvements in en2x translation, suggesting enhanced multilingual semantic coherence.
- Direct Preference Optimization (DPO) demonstrates superior generalization and scalability compared to supervised fine-tuning (SFT) when applied with carefully curated synthetic preference data for translation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.