Research map / Architectures and efficiency / Attention and transformers
Non-autoregressive Machine Translation with Disentangled Context Transformer
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 parallel easy-first inference instead of mask-predict (ablation) | DisCo Transformer · WMT14 EN-DE and WMT16 EN-RO test | BLEU: 26.73 -> 27.34; 30.75 -> 31.31; 33.02 -> 33.22; 33.27 -> 33.25 | Mixed |
| use parallel easy-first inference instead of mask-predict (ablation) | DisCo Transformer · WMT17 EN-ZH test | BLEU: 34.51 -> 34.63; 23.68 -> 23.83 | Helped |
| use parallel easy-first inference instead of mask-predict (ablation) | DisCo Transformer · WMT14 EN-FR en→fr test | BLEU: 40.27 -> 40.66 | Helped |
| train in easy-first order rather than random context sampling (ablation) | DisCo Transformer · WMT14 EN-DE en→de dev | BLEU: 25.60 -> 24.76 | Hurt |
| train in easy-first order rather than random context sampling (ablation) | DisCo Transformer · WMT16 EN-RO ro→en dev | BLEU: 34.97 -> 34.96 | No clear effect |
| use easy-first ordering rather than left-to-right ordering (ablation) | DisCo inference · WMT14 EN-DE en→de dev | BLEU: 21.25 -> 25.60 | Helped |
| use easy-first ordering rather than right-to-left ordering (ablation) | DisCo inference · WMT14 EN-DE en→de dev | BLEU: 20.75 -> 25.60 | Helped |
| use easy-first contexts instead of all-but-itself contexts (ablation) | DisCo inference · WMT14 EN-DE en→de dev | BLEU: 20.72 -> 25.60 | Helped |
| apply sequence-level knowledge distillation (ablation) | CMLM + Mask-Predict · WMT14 EN-DE en→de dev, 4 iterations | BLEU: 22.7 -> 25.5 | Helped |
| apply sequence-level knowledge distillation (ablation) | DisCo + Mask-Predict · WMT14 EN-DE en→de dev, 4 iterations | BLEU: 21.4 -> 24.6 | Helped |
| apply sequence-level knowledge distillation (ablation) | DisCo + Easy-First · WMT16 EN-RO ro→en dev | BLEU: 34.0 -> 35.0 | Helped |
| use contextless keys and values in place of the original decoder keys and values (ablation) | AT Transformer · WMT14 EN-DE en→de test, raw | BLEU: 26.85 -> 27.09 | Helped |
Closest papers by meaning
- Non-Autoregressive Neural Machine Translation (ICLR 2018 · similarity 0.89)
- Improving Non-Autoregressive Translation Models Without Distillation (ICLR 2022 · similarity 0.89)
- Non-Autoregressive Neural Machine Translation with Enhanced Decoder Input (AAAI 2019 · similarity 0.88)
- Asynchronous Bidirectional Decoding for Neural Machine Translation (AAAI 2018 · similarity 0.88)
- Non-autoregressive Translation with Layer-Wise Prediction and Deep Supervision (AAAI 2022 · similarity 0.88)
- 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.