Research map / Learning theory and algorithms / Prediction and decision losses
On the Pareto Front of Multilingual Neural Machine Translation
Liang Chen, Shuming Ma, Dongdong Zhang, Furu Wei, Baobao Chang
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 |
|---|---|---|---|
| compute sampling ratios with Double Power Law instead of tuned temperature (T=2) | MNMT base model · 2-direction En->{DE,HI} | average BLEU: 23.4 -> 23.7 | Improved |
| compute sampling ratios with DPL instead of best temperature (T=5) | MNMT base model · 4-direction | average BLEU: 21.8 -> 21.9 | Improved |
| DPL instead of Gradient Vaccine | MNMT base model · 2-direction | average BLEU: 23.1 -> 23.7 | Improved |
| DPL instead of PCGrad | MNMT base model · 4-direction | average BLEU: 19.6 -> 21.9 | Improved |
| optimize DPL for low-resource Hindi only (r_HI=1) instead of arithmetic mean | MNMT base model · 2-direction En->HI | BLEU: 12.0 -> 12.6 | Improved |
| DPL instead of best temperature T=5 | MNMT base model · 4-direction | average BERT-Score: 70.4 -> 70.8 | Improved |
| DPL instead of best temperature T=5 | MNMT base model · 4-direction | average COMET: 49.7 -> 50.2 | Improved |
| DPL instead of best temperature T=2 | MNMT base model · 2-direction | average cross-entropy: 2.00 -> 1.99 | No clear effect |
| increase temperature from T=1 to T=2 (more weight to low-resource) | MNMT base model · 4-direction | average BLEU: 19.7 -> 21.8 | Improved |
| raise temperature from T=5 to T=100 | MNMT base model · 2-direction | average BLEU: 23.1 -> 18.5 | Hurt |
| model size from Base (64M) to Large (317M) | Double Power Law fit · power term beta | estimated beta: 1.21 -> 3.43 | Improved |
| tune gamma term of DPL to -0.25 for X->EN directions (ablation) | Double Power Law fit · X->EN WMT10 low-resource | r^2: 0.86 -> 0.91 | Improved |
Closest papers by meaning
- Scaling Laws for Multilingual Neural Machine Translation (ICML 2023 · similarity 0.90)
- Meta Back-Translation (ICLR 2021 · similarity 0.87)
- Robust Optimization for Multilingual Translation with Imbalanced Data (NeurIPS 2021 · similarity 0.87)
- Towards Making the Most of BERT in Neural Machine Translation (AAAI 2020 · similarity 0.87)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.86)
- EBBS: An Ensemble with Bi-Level Beam Search for Zero-Shot Machine Translation (AAAI 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.
