Research map / Architectures and efficiency / RNNs and normalization
Minimizing the Bag-of-Ngrams Difference for Non-Autoregressive Neural Machine Translation
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 |
|---|---|---|---|
| fine-tune NAT-Base with BoN objective (BoN-FT) (ablation) | NAT-Base (IRNAT) · WMT14 En->De | BLEU: 16.05 -> 19.27 | Helped |
| train jointly with cross-entropy and BoN from scratch instead of only BoN fine-tuning (ablation) | BoN-FT · WMT14 En->De | BLEU: 19.27 -> 20.75 | Helped |
| add BoN fine-tuning after joint training (ablation) | BoN-Joint · WMT14 En->De | BLEU: 20.75 -> 20.90 | Helped |
| use BoN-FT instead of Reinforce-NAT | NAT-Base · WMT14 En->De | BLEU: 19.15 -> 19.27 | Helped |
| use BoN-Joint+FT on IWSLT16 small dataset | NAT-Base · IWSLT16 En->De | BLEU: 24.13 -> 25.72 | Helped |
| use bigram (n=2) BoN instead of unigram n=1 in joint training (ablation) | BoN-Joint · WMT14 En->De dev | BLEU: 17.73 -> 19.28 | Helped |
| vary n-gram size n from 1 to 4 in BoN-FT (ablation) | BoN-FT · WMT14 En->De dev | BLEU: 18.06 / 18.13 / 18.17 / 18.08 | No clear effect |
| replace cross-entropy loss with BoN (n=2) objective as quality proxy (ablation) | NAT · WMT14 En->De dev | Pearson correlation with BLEU: 0.37 -> 0.70 | Helped |
| train with BoN-Joint instead of cross-entropy only (ablation) | NAT-Base · WMT14 En->De dev | repeated tokens removed: 12.3% -> 2.9% | Helped |
| set alpha=0.1 for combining CE and BoN (ablation) | BoN-Joint · WMT14 En->De dev | BLEU | Helped |
Closest papers by meaning
- Non-Autoregressive Neural Machine Translation with Enhanced Decoder Input (AAAI 2019 · similarity 0.91)
- Non-Autoregressive Neural Machine Translation (ICLR 2018 · similarity 0.91)
- Latent-Variable Non-Autoregressive Neural Machine Translation with Deterministic Inference Using a Delta Posterior (AAAI 2020 · similarity 0.90)
- Guiding Non-Autoregressive Neural Machine Translation Decoding with Reordering Information (AAAI 2021 · similarity 0.90)
- Non-Autoregressive Machine Translation with Auxiliary Regularization (AAAI 2019 · similarity 0.90)
- AEQA-NAT : Adaptive End-to-end Quantization Alignment Training Framework for Non-autoregressive Machine Translation (ICML 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.