Microsoft’s Language Translation AI has Reached Human Levels of Accuracy

Pranav Dar 17 Mar, 2018 • 2 min read

Overview

  • Microsoft’s system can translate sentences in Chinese to English with human accuracy
  • The model was trained on around 2000 sentences in a set of news stories
  • Two methods, Dual Learning and Deliberation Networks, were used to improve the accuracy and quality
  • Yet to be tested on real-time news so expectations should be tempered

 

Introduction

Even with the advances in the  Natural Language Processing field, there have always been nagging doubts about the quality and accuracy of translations from one language to another. Take Google’s translation, for example. While it has steadily improved over the years, you still see a few things grammatically wrong with complex sentences.

                           Source: Wikipedia

To bridge that gap, Microsoft claims it has developed a system that can translate from Chinese to English with the quality and accuracy of humans. The researchers behind this system developed it by training the model on a set of news stories called newstest2017.

In order to ensure that they results of the translations were precise, Microsoft hired external bilingual evaluators to compare the results of the machine’s translations with two independently produced human translations.

The researchers used two methods to develop the AI:

  • Dual Learning: Each time they ran a sentence through the system to translate it from Chinese to English, the team also translated it from English to Chinese. This allowed the system to train and learn from it’s own mistakes.
  • Deliberation Networks: The system was taught to repeat the process of translating the same sentence again and again, refining and improving the responses each time.

Two new techniques were also developed during the training phase to further improve the accuracy of the model.

  • Joint Training: This was used to iteratively boost the Chinese to English and English-to-Chinese translations.
  • Agreement Regularization: According to Microsoft, “with this method, the translation can be generated by having the system read from left to right or from right to left. If these two translation techniques generate the same translation, the result is considered more trustworthy than if they don’t get the same results”.

To understand the mathematics behind the system, you can view Microsoft’s official research paper here.

 

Our take on this

This is quite a huge breakthrough in NLP. But caution should be taken at this stage. This research was conducted on a set of old news stories and as of today, has not yet been tested on real-time news stories. It’s applications could go beyond just translations.

If you’re interested in NLP, I encourage you to check out the research paper which lists how the team went about developing the deep neural network behind this system.

 

Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!

Also, go ahead and participate in our Hackathons, including the DataHack Premier League and Lord of the Machines!

 

Pranav Dar 17 Mar 2018

Senior Editor at Analytics Vidhya. Data visualization practitioner who loves reading and delving deeper into the data science and machine learning arts. Always looking for new ways to improve processes using ML and AI.

Frequently Asked Questions

Lorem ipsum dolor sit amet, consectetur adipiscing elit,

Responses From Readers

Clear