Research map / Architectures and efficiency / RNNs and normalization
Duplex Sequence-to-Sequence Learning for Reversible Machine Translation
Zaixiang Zheng, Hao Zhou, Shujian Huang, Jiajun Chen, Jingjing Xu, Lei 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 duplex (bidirectional with reversible model) instead of simplex single direction (ablation) | REDER · WMT14 En-De | BLEU: 26.20 -> 27.50 | Improved |
| use multitask learning for bidirectional translation instead of duplex (ablation) | REDER · WMT14 En-De | BLEU: 26.20 -> 25.58 | Hurt |
| add multitask learning to AT Transformer-base | Transformer-base · WMT14 En-De | BLEU: 27.60 -> 27.06 | Hurt |
| add back-translation without monolingual data | AT Transformer-base · WMT14 En-De | BLEU: 27.60 -> 27.82 | Improved |
| add knowledge distillation (ablation) | NAT without CTC · WMT14 En-De | BLEU: 11.40 -> 19.50 | Improved |
| add CTC (ablation) | KD NAT · WMT14 En-De | BLEU: 19.50 -> 25.01 | Improved |
| add relative self attention (ablation) | KD+CTC REDER · WMT14 En-De | BLEU: 25.01 -> 25.55 | Improved |
| add reversible architecture (revnet) (ablation) | KD+CTC+R-SAN · WMT14 En-De | BLEU: 25.55 -> 26.20 | Improved |
| add forward-backward agreement and cycle consistency losses (ablation) | REDER · WMT14 En-De | BLEU: 26.20 -> 26.89 | Improved |
| add AT reranking to beam search decoding (ablation) | REDER · WMT14 En-De | BLEU: 26.90 -> 27.50 | Improved |
| use CTC beam search instead of greedy decoding (ablation) | REDER · WMT14 En-De | BLEU: 26.89 -> 26.90 | No clear effect |
| train on separate per-direction KD data instead of raw data (ablation) | REDER · WMT14 En-De | BLEU: 17.85 -> 27.50 | Improved |
Closest papers by meaning
- Sequence to Better Sequence: Continuous Revision of Combinatorial Structures (ICML 2017 · similarity 0.85)
- Universal Transformers (ICLR 2019 · similarity 0.84)
- Non-autoregressive Machine Translation with Disentangled Context Transformer (ICML 2020 · similarity 0.84)
- Controlling Neural Machine Translation Formality with Synthetic Supervision (AAAI 2020 · similarity 0.84)
- High-Fidelity Simultaneous Speech-To-Speech Translation (ICML 2025 · similarity 0.83)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 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.
