International Conference on Learning Representations · 2018 · 337 citations · 0 references
ComputingComputer ScienceTranslation StudiesComputational LinguisticsNatural Language ProcessingTranslation TheoryDeep LearningLanguage ModelingLanguage ModelsMachine TranslationTranslation TechnologyWord EmbeddingsCharacter-based EmbeddingsConvolutional Neural NetworksModel Robustness AnalysisModel Robustness AssessmentNeural Language ModelsNeural Machine TranslationRobustness
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.