Synthetic and Natural Noise Both Break Neural Machine Translation
This addresses a critical reliability issue for machine translation systems in real-world applications, but it is incremental as it builds on existing robustness techniques.
The paper tackled the brittleness of character-based neural machine translation models when handling noisy data, finding that state-of-the-art models fail on moderately noisy texts that humans can comprehend, and showed that a character convolutional neural network model can learn representations robust to multiple noise types.
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.