Source-linked AI summary
Synthetic and Natural Noise Both Break Neural Machine Translation
Yonatan Belinkov, Yonatan Bisk
TL;DR
Character-based NMT models address vocabulary and morphology but are brittle on noisy input that humans understand. The paper tests natural and synthetic noise, evaluates structure-invariant representations and noisy-data training, and finds ensemble-trained charCNN representations robust to multiple noise types while natural noise remains difficult.
Problem
Character-based NMT models provide vocabulary and morphological benefits but can be brittle when trained and tested on clean data.
Method
The paper tests NMT systems against natural and synthetic noise and explores structure-invariant representations alongside robust training on noisy data.
Results
The study finds that state-of-the-art systems fail on noisy texts humans can understand, while a character CNN trained across noise types addresses multiple errors.
Takeaways & Limitations
Robust character-CNN representations can improve handling of multiple trained noise types, but existing models do not easily capture the richness of natural human errors.
Takeaways & Limitations
Natural-noise corpora are limited, and the authors state that more work is needed to immunize NMT models against natural noise.
Abstract
from arXiv · showhide
Character-based neural machine translation (NMT) models alleviate out-of-vocabulary issues, learn morphology, and move us closer to completely end-to-end translation systems. Unfortunately, they are also very brittle and easily falter when presented with noisy data. In this paper, we confront NMT models with synthetic and natural sources of noise. We find that state-of-the-art models fail to translate even moderately noisy texts that humans have no trouble comprehending. We explore two approaches to increase model robustness: structure-invariant word representations and robust training on noisy texts. We find that a model based on a character convolutional neural network is able to simultaneously learn representations robust to multiple kinds of noise.
1 INTRODUCTION
Character-based NMT models offer morphological and vocabulary advantages but remain brittle on noisy text, unlike humans. The paper investigates structure-invariant representations and noisy-data training to improve robustness.
- Humans can often comprehend typos, misspellings, omitted letters, and reordered letters, although the mechanisms and limits remain unknown.
- NMT systems can fail dramatically on noisy input: Google Translate produced an unintelligible translation of a German scrambled-text example.
- Character and sub-word NMT models learn morphological information and generalize to unseen words and conjugations.
- Training and testing on clean data leaves NMT models brittle and potentially unsuitable for broad deployment.
- The paper explores structure-invariant representations and robust training on noisy data, finding that an ensemble-trained character CNN handles multiple noise types.
2 ADVERSARIAL EXAMPLES
Adversarial examples expose how small input changes can cause dramatic model failures, motivating training methods that improve robustness to such perturbations.
- Small input changes can cause dramatic failures in deep learning models, including human-indistinguishable image changes that trigger misclassification.
- Recent NLP work studies white-box and black-box adversarial examples for text classification, evaluation, morphological tagging, and machine translation.
- Adversarial training includes adversarial examples in training data to improve performance on such examples at test time.
- Robustness from adversarial training is sensitive to the example types seen during training, while ensemble training can broaden robustness.
- Noise training can regularize neural networks and expose models to samples outside the training distribution.
3 MT SYSTEMS
The study compares NMT systems with character information at fully character-level, sub-word, and character-CNN word-representation levels.
- The experiments use three NMT systems that access character information at different levels.
- The char2char system is an attention-based sequence-to-sequence model trained from characters to characters with convolutional, highway, recurrent, and decoder layers.
- Nematus is a sequence-to-sequence system operating primarily on byte-pair-encoded sub-word units.
- The authors train an attentional sequence-to-sequence model whose charCNN learns character-dependent word representations while retaining word-level structure.
4 DATA
The paper evaluates clean and noisy source text using natural errors harvested from correction corpora and four synthetic noise processes, including letter scrambling and keyboard typos.
- 4.1 MT DATA: The experiments use the TED talks parallel corpus, official IWSLT 2016 splits, and Moses tokenization for testing and charCNN training.
- 4.2 NATURAL NOISE: Natural noise is approximated by harvesting typos and misspellings from available edit corpora into a lookup table of single-word replacements.
- 4.2 NATURAL NOISE: The natural-error sources cover French Wikipedia edits, small German revision and learner corpora, and diverse errors in Czech non-native essays.
- 4.2 NATURAL NOISE: Collected errors replace source words when available, with uniformly sampled alternatives and unchanged words otherwise; German and Czech data can affect up to half the corpus words.
- 4.2.2 SYNTHETIC NOISE: Synthetic noise includes Swap, Middle Random, Fully Random, and Keyboard Typo transformations.
- 4.2.2 SYNTHETIC NOISE: Swap exchanges two internal letters, Middle Random shuffles internal letters, Fully Random shuffles every letter, and Keyboard Typo substitutes an adjacent key.
5 FAILURES TO TRANSLATE NOISY TEXTS
Models trained on clean text suffer substantial translation degradation on both synthetic and natural noise, despite humans understanding moderately scrambled text. Spell checking can improve French and German results, but natural noise remains difficult to address.
- All models suffer significant BLEU drops when clean-trained systems are tested on noisy texts, with random scrambling producing the lowest scores.
- Humans can understand scrambled texts across languages, while state-of-the-art char2char and Nematus systems fail on the same German example.
- Google Translate’s spell checking often yields a single correction in French and German, corresponding to roughly +5 or more BLEU.
- Even with spell checking, performance remains substantially below vanilla-text results, suggesting natural noise is not easily handled by existing tools.
6 DEALING WITH NOISE
The paper tests structure-invariant representations and noisy-data training as routes to robust translation. MeanChar handles scrambling but not other noise consistently, while charCNN training benefits from noise-specific and broader mixed-noise exposure.
- 6.1 STRUCTURE INVARIANT REPRESENTATIONS: MeanChar is insensitive to scrambling by construction, but remains sensitive to keyboard and natural noise.
- 6.1 STRUCTURE INVARIANT REPRESENTATIONS: MeanChar performs much better than charCNN on scrambled texts, although it trails charCNN by about 7 BLEU points on vanilla French and German.
- 6.2 BLACK-BOX ADVERSARIAL TRAINING: Training meanChar on noisy text improves some matching conditions but does not consistently increase robustness across noise types or languages.
- 6.2 BLACK-BOX ADVERSARIAL TRAINING: CharCNN models trained on a specific noise type perform well on that same type at test time while retaining fairly good quality on vanilla texts.
- 6.2 BLACK-BOX ADVERSARIAL TRAINING: Random-noise training transfers to Swap and Mid scrambling, but models trained on one broad noise class generally do not perform well on the others.
- 6.2 BLACK-BOX ADVERSARIAL TRAINING: Only models trained on natural noise can reasonably translate natural noise at test time, unlike humans who decipher random letter orderings without explicit training.
7 ANALYSIS
The analysis links charCNN robustness to filters learning different responses across noise conditions, while natural noise differs substantially from the synthetic corruptions used in training.
- The authors analyze charCNN filter-weight variance across 1,000 filters and 25 character-embedding dimensions under Rand, Key, Nat, and mixed-noise training.
- Rand-trained models have much smaller weight variances than Key, Nat, and mixed-noise models because random scrambling provides no patterns for filters to detect.
- Mixed-noise training produces intermediate variance as the ensemble captures both Rand’s uniform relationships and Key and Nat’s diverse patterns.
- The variance of variances is smallest for Rand, largest for Key and Nat, and intermediate for mixed noise, indicating greater filter differentiation under Key and Nat.
- 7.2 RICHNESS OF NATURAL NOISE: In a small German sample of about 40 natural-noise examples, phonological phenomena account for 34% and character omissions for 32%.
- 7.2 RICHNESS OF NATURAL NOISE: Synthetic-noise training does not capture common phonological and omission errors, helping explain its weak performance on natural noise.
8 CONCLUSION
The authors find character-based NMT brittle under both natural and synthetic noise, while a character CNN with ensemble training handles multiple trained error types. Natural human errors remain difficult to capture, motivating more realistic noise generation and further robustness work.
- Character-based NMT models break under both natural and synthetic noise.
- A structure-invariant word representation and ensemble training on adversarial examples improve robustness across multiple error types.The character-based CNN learns to address multiple error types seen during training.
- Natural human errors have rich characteristics that existing models cannot easily capture.
- Future work may use phonetic and syntactic structure to generate more realistic synthetic noise.
- Limited corpora of natural noise constrain efforts to immunize NMT models against natural noise.The authors also suggest architectures that could be robust without seeing noise during training.
N. Narodytska and S. Kasiviswanathan. Simple Black-Box Adversarial Attacks on Deep Neural
This passage set consists of bibliographic entries covering adversarial attacks, human word recognition, morphological processing, datasets, and neural machine translation. It provides related-work references rather than a substantive argument or result.
- Several entries cite work on adversarial examples and attacks against machine-learning systems.
- The references include studies of human word and speech recognition under altered or noisy inputs.
- Other entries cover neural machine translation, subword units, morphology, segmentation, and learner or grammatical-error datasets.