--- title: "Towards Robust Neural Machine Translation — Cheng 2018" description: "The authors tested this by perturbing one word per sentence in a test set and found that most… 0.64 BLEU. Cited by 68 papers here." authors: ["Yong Cheng", "Zhaopeng Tu", "Fandong Meng", "Junjie Zhai", "Yang Liu"] year: 2018 venue: "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)" source: "https://aclanthology.org/P18-1163/" doi: "10.18653/v1/p18-1163" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/cheng-2018-robust-neural-machine-translation/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/cheng-2018-robust-neural-machine-translation/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2018](https://md.research.elamur.ai/free/en/papers/2018/page/3/) / Towards Robust Neural Machine Translation # Towards Robust Neural Machine Translation **Authors:** [Yong Cheng](https://md.research.elamur.ai/free/en/authors/yong-cheng/), [Zhaopeng Tu](https://md.research.elamur.ai/free/en/authors/zhaopeng-tu/), [Fandong Meng](https://md.research.elamur.ai/free/en/authors/fandong-meng/), Junjie Zhai and [Yang Liu](https://md.research.elamur.ai/free/en/authors/yang-liu/) · **Year:** 2018 · **Venue:** [Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2018](https://md.research.elamur.ai/free/en/venues/acl/2018/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/P18-1163/) · [PDF](https://aclanthology.org/P18-1163.pdf) · [DOI 10.18653/v1/p18-1163](https://doi.org/10.18653/v1/p18-1163) · [OpenAlex](https://openalex.org/W2964048171) **Citations:** 169 citations (OpenAlex, 2026-10-14) · [68](https://md.research.elamur.ai/free/en/papers/cheng-2018-robust-neural-machine-translation/cited-by/) in our library, of the 141,840 papers whose bibliographies we read ## Summary Neural machine translation systems are fragile: replacing one source word with a synonym can drastically change the output. The authors tested this by perturbing one word per sentence in a test set and found that most translations shifted. Their remedy is adversarial stability training, an architecture-independent training scheme that pushes both encoder and decoder to treat a sentence and a slightly noised version of it the same way. Noisy inputs are built in two ways. The lexical method swaps sampled words for vocabulary items with similar embeddings. The feature method adds Gaussian noise to word embeddings. The objective combines three losses: the usual translation loss, an adversarial invariance loss in which a discriminator tries to separate encoder states of clean and perturbed inputs, and a loss for translating the perturbed source into the original target. Experiments covered Chinese-English, English-German and English-French. The method beat strong baselines, outperformed the RNN-based GNMT, and roughly matched the CNN-based system of Gehring et al. (2017). This matters for noisy real-world input, such as ASR output in simultaneous translation or text with typos. ## Key points - Baseline sensitivity: after replacing a single word with a synonym, 69.74% of test translations changed. BLEU between translations of original and perturbed inputs was only 79.01. - Two perturbation variants are proposed: - AST lexical samples replacements by embedding cosine similarity. - AST feature injects Gaussian noise into embeddings. - The framework also accepts task-specific noise, such as homonyms for speech-based pipelines. - Ablation on Chinese-English (AST lexical): - Dropping the adversarial loss cost 0.64 BLEU. - Dropping L_noisy cost 1.66 BLEU. - Using L_noisy alone still gave a 0.88 BLEU gain over the baseline. - L_inv settles near 0.68 after about 100K iterations. This means the discriminator outputs about 0.5 for both clean and perturbed samples, so the encoder treats them almost identically. - Training is less stable than MLE. BLEU first drops by up to about 7 points before recovering, and the loss curves oscillate. The method needs no extra monolingual or bilingual corpora. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/) — University, Beijing, China · 137 works · 83,262 citations - [Tencent](https://md.research.elamur.ai/free/en/companies/tencent/) — Company, Shenzhen, China · 66 works · 17,945 citations - [State Key Laboratory of Intelligent Technology and Systems](https://md.research.elamur.ai/free/en/labs/state-key-laboratory-of-intelligent-technology-and-systems/) — Research institute, Beijing, China · 21 works · 9,641 citations - [Beijing National Research Center for Information Science and Technology](https://md.research.elamur.ai/free/en/labs/beijing-national-research-center-for-information-science-and-technology/) — Research institute, Beijing, China · 21 works · 5,843 citations Also on the byline: [Beijing Advanced Innovation Center for Language Resources](https://ror.org/050djy369). 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works 🧭 Topics (1): [🔁 Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) · 958 works ## Cite Cite as: Yong Cheng, Zhaopeng Tu, Fandong Meng, Junjie Zhai and Yang Liu. 2018. Towards Robust Neural Machine Translation. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{cheng2018robust, title = {{Towards Robust Neural Machine Translation}}, author = {Yong Cheng, Zhaopeng Tu, Fandong Meng, Junjie Zhai and Yang Liu}, booktitle = {Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, year = {2018}, doi = {10.18653/v1/p18-1163}, url = {https://doi.org/10.18653/v1/p18-1163} } ``` - [← previous Comparing Automatic and Human Evaluation of Local Explanations for Text Classification](https://md.research.elamur.ai/free/en/papers/nguyen-2018-comparing-automatic-human-evaluation-local/) - [next → Document Context Neural Machine Translation with Memory Networks](https://md.research.elamur.ai/free/en/papers/maruf-2018-document-context-neural-machine-translation/) - [↑ back to papers of 2018](https://md.research.elamur.ai/free/en/papers/2018/page/3/) *Updated 24 October 2026*