Research map / Large language models / Language, safety and interpretability
Controlling Neural Machine Translation Formality with Synthetic Supervision
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 |
|---|---|---|---|
| Attach style tags to both source and target instead of source only | Formality control model · GYAFC I→F and formality-preservation tasks | BLEU: 72.16 ± 0.34 -> 72.29 ± 0.23; 66.87 ± 0.58 -> 67.81 ± 0.41; 78.78 ± 0.37 -> 79.34 ± 0.55 | Helped |
| Block source-tag visibility from the encoder (ablation) | Tag-Src · GYAFC I→I and F→F | BLEU: 66.87 ± 0.58 -> 65.46 ± 0.29; 78.78 ± 0.37 -> 76.72 ± 0.39 | Hurt |
| Infer target style online from each parallel translation pair | Multi-Task FSMT · WMT newstest2014 | LeD: 10.89 -> 14.53 | Helped |
| Infer target style online from each parallel translation pair | Multi-Task FSMT · WMT newstest2014 | PoD: 7.76 -> 12.58 | Helped |
| Infer target style online from each parallel translation pair | Multi-Task FSMT · MSLT | LeD: 11.97 -> 14.52 | Helped |
| Infer target style online from each parallel translation pair | Multi-Task FSMT · MSLT | PoD: 1.41 -> 2.19 | Helped |
| Generate synthetic target sequences rather than infer the reference translation's formality (ablation) | Multi-Task FSMT · WMT newstest2014 | PoD: 7.76 -> 7.25 (Online Target Inference is reported as higher than Multi-Task in positional discrepancy, but this table va | Helped |
| Generate synthetic target sequences rather than infer the reference translation's formality (ablation) | Multi-Task FSMT · MSLT | PoD: 1.41 -> 1.63 | Helped |
| Use online style inference rather than Multi-Task supervision | Multi-Task FSMT · WMT and MSLT human evaluation | formality difference: 151:80:52 (informal); 153:84:61 (formal) | Helped |
| Use online style inference rather than Multi-Task supervision | Multi-Task FSMT · WMT and MSLT human evaluation | meaning preservation: 205:217:155 | Helped |
| Use Factor-Concat style conditioning instead of Tag-Src-Tgt (ablation) | Formality control model · GYAFC four-way formality rewriting | BLEU: +0.18; 0.00; -0.78; +0.46 | Mixed |
| Inject style only into the decoder using Pred-Sum (ablation) | Formality control model · GYAFC I→I | BLEU: 67.81 ± 0.41 -> 66.15 ± 0.41 | Hurt |
Closest papers by meaning
- Empirical Regularization for Synthetic Sentence Pairs in Unsupervised Neural Machine Translation (AAAI 2021 · similarity 0.89)
- Generative Neural Machine Translation (NeurIPS 2018 · similarity 0.88)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.88)
- Balancing Quality and Human Involvement: An Effective Approach to Interactive Neural Machine Translation (AAAI 2020 · similarity 0.88)
- Non-Autoregressive Neural Machine Translation (ICLR 2018 · similarity 0.87)
- Mirror-Generative Neural Machine Translation (ICLR 2020 · 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.