Research map / Architectures and efficiency / RNNs and normalization
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 |
|---|---|---|---|
| Train on teacher-generated distilled targets instead of ground-truth targets (ablation) | NAT with uniform source copying · IWSLT16 En→De development | BLEU: 16.51 -> 20.72 | Helped |
| Replace uniform source copying with fertility-based copying (ablation) | NAT trained on ground-truth targets · IWSLT16 En→De development | BLEU: 16.51 -> 18.87 | Helped |
| Replace uniform source copying with fertility-based copying (ablation) | NAT trained on distilled targets · IWSLT16 En→De development | BLEU: 20.72 -> 25.20 | Helped |
| Add positional attention (ablation) | NAT with fertility copying and greedy distillation · IWSLT16 En→De development | BLEU: 24.02 -> 25.20 | Helped |
| Add all three fine-tuning terms together (ablation) | NAT with fertility copying, positional attention, and greedy distillation · IWSLT16 En→De development | BLEU: 25.20 -> 26.52 | Helped |
| Use reinforcement-learning fine-tuning alone (ablation) | NAT · IWSLT16 En→De development | BLEU | Hurt |
| Use the backpropagation fine-tuning term alone (ablation) | NAT · IWSLT16 En→De development | BLEU | Hurt |
| Use positional embeddings alone as decoder inputs (ablation) | NAT decoder · IWSLT16 En→De development | BLEU: \approx 2 | Hurt |
| Increase noisy parallel decoding sample count from 10 to 100 | NAT with fine-tuning · WMT16 En→Ro newstest2016 | BLEU: 29.02 -> 29.79 | Helped |
| Add fine-tuning | NAT · IWSLT16 En→De development | BLEU: 25.20 -> 26.52 | Helped |
| Use NPD with 100 fertility samples rather than argmax decoding | NAT with fine-tuning · WMT16 En→Ro newstest2016 | BLEU: 27.29 -> 29.79 | Helped |
| Increase NPD sample count from 10 to 100 | NAT with fine-tuning · WMT14 En→De newstest2014 | BLEU: 18.66 -> 19.17 | Helped |
Closest papers by meaning
- Latent-Variable Non-Autoregressive Neural Machine Translation with Deterministic Inference Using a Delta Posterior (AAAI 2020 · similarity 0.93)
- Deep Encoder, Shallow Decoder: Reevaluating Non-autoregressive Machine Translation (ICLR 2021 · similarity 0.92)
- Improving Non-Autoregressive Translation Models Without Distillation (ICLR 2022 · similarity 0.91)
- Non-Autoregressive Neural Machine Translation with Enhanced Decoder Input (AAAI 2019 · similarity 0.91)
- Non-autoregressive Translation with Layer-Wise Prediction and Deep Supervision (AAAI 2022 · similarity 0.91)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 2025 · similarity 0.88)
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.