New model to overcome weakness of Google Neural Network machine translation

Google’s neural machine translation has a problem in that if the sentence before translation is subtly modified, the sentence after translation will change significantly. In order to overcome this weakness, Google newly introduced a translation model that introduces an algorithm that confuses humans by placing an unidentifiable noise signal on the image (Adversarial Examples).

Neural network machine translation using the Google Transformer model requires explicit language rules based on deep neural networks and translates into an end-to-end parallel corpus. However, as previously stated, neural network machine translation has a weakness in that it is sensitive to subtle changes in input information. If only one word in the text is changed to a synonym, there is a possibility that the translation will be completely different.

Some companies and organizations say that neural network machine translation cannot be integrated into the system because of its lack of robustness. It was also pointed out that Wikipedia’s own credibility was compromised as a result of posting many texts to machine translation.

Google is doing research to solve this problem. A paper published in June is one of these methods. It introduces an algorithm called Adversarial Examples that confuses the translation model by putting noise signals that cannot be identified by humans. This technology is inspired by the hostile generation network GAN, and does not rely on a discriminator to determine the authenticity, but introduces hostile cases into learning and diversifies and expands the training set.

As a result of benchmarking translations in the Chinese-English and English-German combinations, the development team revealed that the BLEU score improved by 2.8 and 1.6 points, respectively, compared to the existing transformer model.

This research result can be said to be meaningful in that it showed the possibility of overcoming the weakness of the existing neural network machine translation, such as lack of robustness. The new model shows high performance even compared to competing models, so it is expected that this translation model will be useful for downstream work in the future. Related information can be found here .