Amazon at WMT: Improving Machine Translation with User Feedback

Machine translation is one of the most exciting and important applications of machine learning.  It’s a widely researched topic, both in the academic community and within major technology companies, such as Amazon.  At Amazon, we use machine translation to do things like document the same products in multiple languages.  This helps us to offer additional language options to our customers. For example, customers can now switch their language preference to Spanish on Amazon.com  or to English, Dutch, Polish, or Turkish on Amazon.de.

Since 2006, the important annual event Workshop on Machine Translation invites participants to submit machine translation systems for competitive ranking in a number of categories.  This year Amazon, in collaboration with Germany’s Heidelberg University, is hosting a new competition for machine translation systems that adapt well to simulated customer feedback; in other words, systems that are able to correct their mistakes by learning from a stream of translation assessments.  The results will be presented at this year’s WMT Conference in Copenhagen.

To support this competition, we’re using many AWS services, including Amazon API Gateway to host our service front end and provide an SDK, AWS Lambda to perform backend computations, and Amazon DynamoDB to store the state of experiments and our training data. Finally, we’re using Amazon CloudWatch to monitor the service and Amazon SNS to notify us when an alarm is triggered.

To learn more about the competition, see the WMT17 Bandit Learning Task page.

If you’re a researcher in the machine translation field and are interested in the competition, there’s still time to sign up.  Contact us at mt-shared-task@amazon.com.