Research map / Label-efficient and robust learning / Recognition and long-tail
Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translation
Víctor M. Sánchez-Cartagena, Miquel Esplà-Gomis, Juan Antonio Pérez-Ortiz, Felipe Sánchez-Martínez
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 |
|---|---|---|---|
| add MaTiLDA (reverse+swap+replace) synthetic non-fluent data to the baseline | Transformer base · low-resource in-domain, 10 directions | BLEU: 1.1 (hsb--de) to 4.1 (de--rm) BLEU points, 2.1 on average | Improved |
| use MaTiLDA instead of the best reference DA system (SeqMix etc.) | Transformer base · low-resource in-domain | BLEU: 0.4 to 2.5 BLEU points; 1.1 on average | Improved |
| add the replace transformation alone (ablation) | MaTiLDA · de-rm low-resource | BLEU: 49.3 -> 53.3 | Improved |
| add the source (copy source sentence) transformation alone (ablation) | MaTiLDA · hsb-de low-resource | BLEU: 49.1 -> 48.4 | Hurt |
| add the mono transformation alone (ablation) | MaTiLDA · en-de low-resource | BLEU: 25.4 -> 25.2 | Hurt |
| combine reverse+swap+replace instead of replace alone (ablation) | MaTiLDA · rm-de low-resource | BLEU: 52.7 -> 54.1 | Improved |
| combine transformations (rev+sw+repl) instead of single swap (ablation) | MaTiLDA · hsb-de low-resource | BLEU: 50.2 -> 49.7 | Hurt |
| add MaTiLDA on top of back-translation | back-translation baseline · de-en low-resource with back-translated data | BLEU: 31.4 -> 32.8 | Improved |
| add MaTiLDA on top of back-translation | back-translation baseline · rm-de low-resource with back-translated data | BLEU: 49.1 -> 51.3 | Improved |
| add MaTiLDA (swap+unk+replace) in high-resource conditions | Transformer base · en-de high-resource | BLEU: 24.3 -> 25.8 | Improved |
| add MaTiLDA (swap+unk+replace) in high-resource conditions | Transformer base · en-ro high-resource | BLEU: 23.3 -> 24.3 | Improved |
| add MaTiLDA for out-of-domain translation (low-resource) | Transformer base · de-en IT out-of-domain low-resource | BLEU: 4.5 -> 13.3 | Improved |
Closest papers by meaning
- Controlling Neural Machine Translation Formality with Synthetic Supervision (AAAI 2020 · similarity 0.90)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.88)
- Non-Autoregressive Neural Machine Translation (ICLR 2018 · similarity 0.88)
- Minimizing the Bag-of-Ngrams Difference for Non-Autoregressive Neural Machine Translation (AAAI 2020 · similarity 0.88)
- Non-Autoregressive Neural Machine Translation with Enhanced Decoder Input (AAAI 2019 · similarity 0.88)
- Meta Back-Translation (ICLR 2021 · similarity 0.88)
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.
