Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Neural Machine Translation with SMT as Posterior Regularization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use SMT as posterior regularization in joint EM training with NMT | unsupervised NMT · fr-en | BLEU: 27.68 -> 28.79 | Helped |
| use SMT as posterior regularization in joint EM training with NMT | unsupervised NMT · en-fr | BLEU: 28.11 -> 29.21 | Helped |
| use SMT as posterior regularization in joint EM training with NMT | unsupervised NMT · de-en newstest 2014 | BLEU: 10.21 -> 20.04 | Helped |
| use SMT as posterior regularization in joint EM training with NMT | unsupervised NMT · en-de newstest 2014 | BLEU: 6.89 -> 16.43 | Helped |
| use SMT as posterior regularization in joint EM training with NMT | unsupervised NMT · de-en newstest 2016 | BLEU: 25.19 -> 25.92 | Helped |
| use SMT as posterior regularization in joint EM training with NMT | unsupervised NMT · en-de newstest 2016 | BLEU: 20.23 -> 21.07 | Helped |
| add R2L regularization after the final training iteration (ablation) | Our Method · English-French and English-German test sets | BLEU: 28.79 -> 28.92; 29.21 -> 29.53; 20.04 -> 20.43; 16.43 -> 16.97; 25.92 -> 26.32; 21.07 -> 21.65 | Helped |
| replace SMT0 warm-up with word-by-word pseudo parallel initialization (ablation) | Our Method initialization · fr-en newstest 2014 | BLEU: 24.06 -> 12.29 | Hurt |
| replace SMT0 warm-up with word-by-word pseudo parallel initialization (ablation) | Our Method initialization · en-fr newstest 2014 | BLEU: 24.82 -> 12.46 | Hurt |
| replace SMT0 warm-up with word-by-word pseudo parallel initialization (ablation) | Our Method initialization · de-en newstest 2014 | BLEU: 16.29 -> 7.32 | Hurt |
| replace SMT0 warm-up with word-by-word pseudo parallel initialization (ablation) | Our Method initialization · en-de newstest 2014 | BLEU: 12.88 -> 4.81 | Hurt |
| add the SMT E-step between NMT training steps | Our Method · newstest 2014, four translation directions | BLEU: +2.06 | Helped |
Closest papers by meaning
- Neural Machine Translation Advised by Statistical Machine Translation (AAAI 2017 · similarity 0.91)
- Unsupervised Neural Machine Translation (ICLR 2018 · similarity 0.91)
- Improved Neural Machine Translation with SMT Features (AAAI 2016 · similarity 0.91)
- Joint Training for Neural Machine Translation Models with Monolingual Data (AAAI 2018 · similarity 0.90)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.87)
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.87)
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.