Research map / Label-efficient and robust learning / Recognition and long-tail
Cross-model Back-translated Distillation for Unsupervised Machine Translation
Xuan-Phi Nguyen, Shafiq Joty, Thanh-Tung Nguyen, Kui Wu, Ai Ti Aw
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 |
|---|---|---|---|
| Add cross-model back-translated distillation to XLM | XLM · Base WMT, En-Fr, Fr-En, En-De, De-En, En-Ro, Ro-En | BLEU: 33.0 -> 35.4; 31.5 -> 33.0; 23.9 -> 26.1; 29.3 -> 31.5; 30.6 -> 32.2; 27.9 -> 29.2 | Improved |
| Add cross-model back-translated distillation to NMT | NMT · Base WMT, En-Fr, Fr-En, En-De, De-En, En-Ro, Ro-En | BLEU: 24.7 -> 26.6; 24.5 -> 25.7; 14.5 -> 16.6; 18.2 -> 20.5; 16.7 -> 18.1; 16.3 -> 17.8 | Improved |
| Add cross-model back-translated distillation to PBSMT | PBSMT · Base WMT, En-Fr, Fr-En, En-De, De-En, En-Ro, Ro-En | BLEU: 17.1 -> 21.6; 16.4 -> 20.6; 10.9 -> 15.0; 13.6 -> 17.7; 10.5 -> 11.3; 11.7 -> 14.5 | Improved |
| Add CBD to NMT | NMT · IWSLT En-Fr and En-De, four translation directions | BLEU: 29.6 -> 31.8; 30.7 -> 31.8; 15.8 -> 18.4; 19.1 -> 21.7 | Improved |
| Use one agent for one-way synthetic data rather than cross-model back-translation (ablation) | CBD · Base WMT En-Fr | BLEU: 26.6 -> 24.5 | Hurt |
| Use two agents for separate forward translations without cross-model back-translation (ablation) | CBD · Base WMT En-Fr | BLEU: 26.6 -> 24.6 | Hurt |
| Use the same agent for forward and backward translation (ablation) | CBD · Base WMT En-Fr | BLEU: 26.6 -> 24.8 | Hurt |
| Increase the number of agents from 2 to 3 (ablation) | GCBD · IWSLT En-Fr, Fr-En, En-De, De-En | BLEU: 31.8 -> 32.8; 31.8 -> 32.1; 18.4 -> 19.2; 21.7 -> 22.2 | Improved |
| Increase the number of agents from 3 to 4 (ablation) | GCBD · IWSLT En-Fr, Fr-En, En-De, De-En | BLEU: 32.8 -> 32.3; 32.1 -> 32.0; 19.2 -> 19.1; 22.2 -> 21.9 | Hurt |
| Replace CBD with ensemble knowledge distillation | Unsupervised MT with NMT agents · Base WMT En-Fr, Fr-En, En-De, De-En, En-Ro, Ro-En | BLEU: CBD: 26.6, 25.7, 16.6, 20.5, 18.1, 17.8; Ens-Distil: 17.3, 20.0, 3.5, 3.7, 1.2, 1.1 | Hurt |
| Use temperature sampling instead of deterministic CBD | XLM · Base WMT En-Fr, Fr-En, En-De, De-En | BLEU: 33.0 -> 33.5; 31.5 -> 32.2; 23.9 -> 24.3; 29.3 -> 30.2 | Mixed |
| Use target noising during iterative back-translation | XLM · Base WMT En-Fr, Fr-En, En-De, De-En | BLEU: 33.0 -> 32.8; 31.5 -> 30.7; 23.9 -> 24.0; 29.3 -> 29.6 | Mixed |
Closest papers by meaning
- Multilingual Neural Machine Translation with Knowledge Distillation (ICLR 2019 · similarity 0.88)
- Cross-lingual Language Model Pretraining (NeurIPS 2019 · similarity 0.88)
- Refining Low-Resource Unsupervised Translation by Language Disentanglement of Multilingual Translation Model (NeurIPS 2022 · similarity 0.88)
- Multi-Level Optimal Transport for Universal Cross-Tokenizer Knowledge Distillation on Language Models (AAAI 2025 · similarity 0.86)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.86)
- Teaching Large Language Models to Translate with Comparison (AAAI 2024 · similarity 0.85)
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.
