Research map / Label-efficient and robust learning / Recognition and long-tail
Refining Low-Resource Unsupervised Translation by Language Disentanglement of Multilingual Translation Model
Xuan-Phi Nguyen, Shafiq Joty, Kui Wu, AiTi 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 stage 1 (language-specific decoder FFN finetuning) (ablation) | CRISS · FLoRes En-Ne | BLEU: 5.5 -> 7.8 | Improved |
| add stage 1 (ablation) | CRISS · FLoRes Gu-En | BLEU: 23.7 -> 26.2 | Improved |
| add stage 2 (separate English) after stage 1 (ablation) | Ours · FLoRes Gu-En | BLEU: 26.2 -> 28.8 | Improved |
| add stage 2 after stage 1 (ablation) | Ours · FLoRes En-Et | BLEU: 17.4 -> 19.7 | Improved |
| add stage 3 after stage 2 (ablation) | Ours · FLoRes En-Lv | BLEU: 16.3 -> 17.9 | Improved |
| add stage 4 after stage 3 in from-English direction (ablation) | Ours · FLoRes En-Gu | BLEU: 16.8 -> 17.5 | Improved |
| add stage 4 after stage 3 in to-English direction (ablation) | Ours · FLoRes Gu-En | BLEU: 29.5 -> 29.1 | No clear effect |
| add stage 4 after stage 3 in to-English direction (ablation) | Ours · FLoRes Kk-En | BLEU: 20.0 -> 19.9 | No clear effect |
| add decoder language-specific FFN layers (ablation) | Ours · FLoRes En-Ne | BLEU: 7.9 -> 8.8 | Improved |
| use language-specific attention instead of FFN in decoder (ablation) | Ours · FLoRes En-Ne | BLEU: 8.8 -> 8.4 | Hurt |
| place language-specific FFN in encoder instead of decoder (ablation) | Ours · FLoRes En-Ne | BLEU: 8.8 -> 6.5 | Hurt |
| add language-specific FFN to encoder as well as decoder (ablation) | Ours · FLoRes En-Ne | BLEU: 8.8 -> 8.9 | No clear effect |
Closest papers by meaning
- Cross-model Back-translated Distillation for Unsupervised Machine Translation (ICML 2021 · similarity 0.88)
- Mufu: Multilingual Fused Learning for Low-Resource Translation with LLM (ICLR 2025 · similarity 0.87)
- Robust Optimization for Multilingual Translation with Imbalanced Data (NeurIPS 2021 · similarity 0.87)
- Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? (ICLR 2025 · similarity 0.86)
- Constrained Decoding for Cross-lingual Label Projection (ICLR 2024 · similarity 0.84)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.84)
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.
