Research map / Large language models / Language, safety and interpretability
Beyond English-Centric Multilingual Machine Translation
Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch and 4 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| mine pairs with the bridge language strategy instead of random 80% pairs | 1.2B multilingual model · validation, all supervised+zero-shot | BLEU avg: 11.9 -> 17.2 | Improved |
| mine with bridge language strategy instead of English-centric plus random pairs | 1.2B multilingual model · validation | BLEU avg: 16.3 -> 17.2 | Improved |
| mine the full language matrix (0% sparse) instead of 60% sparse | bridge language mining · 50 languages | BLEU | No clear effect |
| add Sinkhorn temperature sampling over plain temperature rescaling | 1.2B multilingual model · validation | BLEU all avg: 18.1 -> 18.6 | Improved |
| use temperature rescaling instead of uniform sampling | 1.2B multilingual model · validation low-resource | BLEU: 6.1 -> 10.2 | Improved |
| train on Many-to-Many data instead of English-centric data | 1.2B model · non-English directions | BLEU: 5.7 -> 15.9 | Improved |
| train on Many-to-Many data instead of English-centric with pivoting | 1.2B model · non-English directions | BLEU: 10.4 -> 15.9 | Improved |
| train on Many-to-Many data instead of English-centric data | 1.2B model · directions involving English | BLEU: 31.0 -> 31.2 | No clear effect |
| add backtranslated data for 100 low-quality directions | 1.2B multilingual model · 100 directions with 2-10 BLEU | BLEU | Improved |
| add random re-routing at 20% of training samples (ablation) | language-specific layers · low and mid resource languages | BLEU: about 0.8 BLEU | Improved |
| scale dense model from 1.2B to 12B parameters | dense Transformer · 100 languages validation | BLEU all avg: 17.5 -> 18.6 | Improved |
| add 6 language-specific decoder layers instead of scaling densely to 12B | 1.2B Transformer · 100 languages validation | BLEU all avg: 18.6 -> 19.2 | Improved |
Closest papers by meaning
- Scaling Laws for Multilingual Neural Machine Translation (ICML 2023 · similarity 0.91)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.90)
- A Benchmark for Learning to Translate a New Language from One Grammar Book (ICLR 2024 · similarity 0.87)
- Exploring Polyglot Harmony: On Multilingual Data Allocation for Large Language Models Pretraining (NeurIPS 2025 · similarity 0.87)
- Enhancing Non-English Capabilities of English-Centric Large Language Models Through Deep Supervision Fine-Tuning (AAAI 2025 · similarity 0.87)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
