AI Paper Atlas

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

JMLR 2021 · Large language models › Language, safety and interpretability · approach: Multilingual LLM data and culture · 610 citations (OpenAlex)

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 changedWhereMeasuredEffect
mine pairs with the bridge language strategy instead of random 80% pairs1.2B multilingual model · validation, all supervised+zero-shotBLEU avg: 11.9 -> 17.2Improved
mine with bridge language strategy instead of English-centric plus random pairs1.2B multilingual model · validationBLEU avg: 16.3 -> 17.2Improved
mine the full language matrix (0% sparse) instead of 60% sparsebridge language mining · 50 languagesBLEUNo clear effect
add Sinkhorn temperature sampling over plain temperature rescaling1.2B multilingual model · validationBLEU all avg: 18.1 -> 18.6Improved
use temperature rescaling instead of uniform sampling1.2B multilingual model · validation low-resourceBLEU: 6.1 -> 10.2Improved
train on Many-to-Many data instead of English-centric data1.2B model · non-English directionsBLEU: 5.7 -> 15.9Improved
train on Many-to-Many data instead of English-centric with pivoting1.2B model · non-English directionsBLEU: 10.4 -> 15.9Improved
train on Many-to-Many data instead of English-centric data1.2B model · directions involving EnglishBLEU: 31.0 -> 31.2No clear effect
add backtranslated data for 100 low-quality directions1.2B multilingual model · 100 directions with 2-10 BLEUBLEUImproved
add random re-routing at 20% of training samples (ablation)language-specific layers · low and mid resource languagesBLEU: about 0.8 BLEUImproved
scale dense model from 1.2B to 12B parametersdense Transformer · 100 languages validationBLEU all avg: 17.5 -> 18.6Improved
add 6 language-specific decoder layers instead of scaling densely to 12B1.2B Transformer · 100 languages validationBLEU all avg: 18.6 -> 19.2Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy