Research map / Label-efficient and robust learning / Noisy labels
Synthetic and Natural Noise Both Break 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 |
|---|---|---|---|
| Evaluate clean-trained NMT on noisy source text | charCNN, char2char, and Nematus · German→English and Czech→English; synthetic and natural noise | BLEU: 34.79 -> 1.02 (German charCNN, clean to Rand); 29.65 -> 0.66 (Czech Nematus, clean to Rand) | Hurt |
| Train charCNN on Rand+Key+Nat instead of clean text | charCNN · French→English, natural noise | BLEU: 17.42 -> 31.77 | Helped |
| Train charCNN on Rand+Key+Nat instead of clean text | charCNN · German→English, natural noise | BLEU: 14.02 -> 31.60 | Helped |
| Train charCNN on Rand+Key+Nat instead of clean text | charCNN · Czech→English, natural noise | BLEU: 10.20 -> 18.38 | Helped |
| Train charCNN on a mix of Rand, Key, and Nat rather than one noise type | charCNN · French, German, and Czech→English; average across tested noise conditions | BLEU: best result on average: 35.70 (French), 30.70 (German), 20.10 (Czech) | Mixed |
| Train meanChar on noisy text rather than clean text | meanChar · French, German, and Czech→English; multiple noise types | BLEU | Mixed |
| Use natural-noise training rather than clean training | charCNN · French→English, natural noise | BLEU: 17.42 -> 36.00 | Helped |
| Use natural-noise training rather than clean training | charCNN · German→English, natural noise | BLEU: 14.02 -> 34.59 | Helped |
| Use natural-noise training rather than clean training | charCNN · Czech→English, natural noise | BLEU: 10.20 -> 21.64 | Helped |
| Correct naturally noisy input with Google spelling suggestions | Google Translate · French natural noise | BLEU: 16.7 -> 21.4 | Helped |
| Correct naturally noisy input with Google spelling suggestions | Google Translate · German natural noise | BLEU: 18.6 -> 25.0 | Helped |
| Correct naturally noisy input with Google spelling suggestions | Google Translate · Czech natural noise | BLEU: 12.3 -> 11.2 | Hurt |
Closest papers by meaning
- Analyzing Uncertainty in Neural Machine Translation (ICML 2018 · similarity 0.88)
- Identifying and Controlling Important Neurons in Neural Machine Translation (ICLR 2019 · similarity 0.86)
- Scaling Laws for Neural Machine Translation (ICLR 2022 · similarity 0.86)
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.86)
- Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models (ICLR 2024 · similarity 0.85)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · 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.