Research map / Learning theory and algorithms / Combinatorial and search optimization
Neural Machine Translation with Gumbel-Greedy Decoding
Jiatao Gu, Daniel Jiwoong Im, Victor O.K. Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 GGD-GAN instead of MLE | NMT with greedy decoding · DE-EN, EN-DE, CS-EN, and EN-CS test sets | BLEU: 21.63 -> 22.47; 18.97 -> 19.38; 18.90 -> 20.12; 14.49 -> 15.4 | Improved |
| Train GGD-GAN instead of REINFORCE | NMT with greedy decoding · DE-EN, EN-DE, CS-EN, and EN-CS test sets | BLEU: 22.56 -> 22.47; 19.32 -> 19.38; 19.45 -> 20.12; 15.02 -> 15.4 | Mixed |
| Train GGD-GAN instead of MLE | NMT with beam search (size=5) · DE-EN, EN-DE, CS-EN, and EN-CS test sets | BLEU: 24.46 -> 25.32; 21.33 -> 21.27; 21.2 -> 21.17; 16.2 -> 16.44 | Mixed |
| Jointly train the discriminator with the generator rather than keeping it fixed (ablation) | GGD · Cs-En validation | BLEU: ≈+1.3 and ≈ +0.6 | Improved |
| Add entropy regularization to generator training (ablation) | GGD generator · Cs-En validation | BLEU | Improved |
| Use sampling-based generation instead of greedy generation (ablation) | GGD · Cs-En validation | BLEU | Improved |
| Set Gumbel-Softmax temperature to 0.5 rather than the other tested values (ablation) | GGD · Cs-En validation | BLEU | Improved |
| Use temperature 5 (ablation) | GGD · Cs-En validation | BLEU | Hurt |
| Anneal the Gumbel-Softmax temperature during training (ablation) | GGD · NMT experiments | BLEU | No clear effect |
| Use ST-Gumbel rather than the original ST estimator (ablation) | GGD generator · Greedy decoding and sampling experiments | BLEU | Improved |
Closest papers by meaning
- Neural Machine Translation with Soft Prototype (NeurIPS 2019 · similarity 0.86)
- Decoding with Value Networks for Neural Machine Translation (NeurIPS 2017 · similarity 0.85)
- Generative Neural Machine Translation (NeurIPS 2018 · similarity 0.85)
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.84)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.83)
- EBBS: An Ensemble with Bi-Level Beam Search for Zero-Shot Machine Translation (AAAI 2025 · similarity 0.82)
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.
