Research map / Label-efficient and robust learning / Knowledge distillation
Understanding and Improving Lexical Choice in Non-Autoregressive 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 |
|---|---|---|---|
| Add the combined raw-data prior | Mask-Predict NAT · WMT14 En-De | BLEU: 27.0 -> 27.8 | Helped |
| Add the combined raw-data prior | Levenshtein NAT · WMT16 Ro-En | BLEU: 33.2 -> 33.8 | Helped |
| Add WAD (ablation) | MaskT+KD · En-De, Zh-En, and Ja-En | BLEU: 27.0 -> 27.4; 24.2 -> 24.8; 28.9 -> 29.4 | Helped |
| Add SDD (ablation) | MaskT+KD · En-De, Zh-En, and Ja-En | BLEU: 27.0 -> 27.5; 24.2 -> 24.9; 28.9 -> 29.3 | Helped |
| Combine WAD and SDD instead of using either alone (ablation) | MaskT+KD · En-De, Zh-En, and Ja-En | BLEU: 27.4 / 27.5 -> 27.8; 24.8 / 24.9 -> 25.2; 29.4 / 29.3 -> 29.6 | Helped |
| Expose raw training data through lexical priors | MaskT+KD · En-De, Zh-En, and Ja-En | AoLC / LFT: 76.3 / 68.4 -> 78.1 / 72.4; 72.7 / 61.5 -> 74.0 / 65.0; 75.3 / 66.9 -> 76.6 / 69.1 | Helped |
| Use data mixing instead of lexical priors (ablation) | MaskT+KD · En-De | BLEU: 27.0 -> 27.2 | Helped |
| Use tagged data mixing instead of lexical priors (ablation) | MaskT+KD · En-De | BLEU: 27.0 -> 27.4 | Helped |
| Use decay curriculum instead of lexical priors (ablation) | MaskT+KD · En-De | BLEU: 27.0 -> 27.5 | Helped |
| Use the proposed raw-data prior | MaskT+KD · En-De | BLEU: 27.0 -> 27.8 | Helped |
| Replace GIZA++ alignments with fast-align (ablation) | Raw-data-prior NAT · En-De and Zh-En | BLEU improvement over baseline: +0.6 and +0.7 vs. +0.8 and +1.0 BLEU | Helped |
| Inject noise into alignment distributions at 2% and 5% (ablation) | Raw-data-prior NAT · En-De | BLEU: -0.1 and -0.2 BLEU | Hurt |
Closest papers by meaning
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.86)
- Trans-EnV: A Framework for Evaluating the Linguistic Robustness of LLMs Against English Varieties (NeurIPS 2025 · similarity 0.86)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.85)
- DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation (NeurIPS 2024 · similarity 0.85)
- Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation (NeurIPS 2025 · similarity 0.85)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.85)
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.