MSRA won in eight out of eleven machine translation tasks it undertook as part of the challenge
SINGAPORE - Media OutReach - 22 May 2019 - Microsoft Research Asia (MSRA) has achieved eight top places in the recent machinetranslation challenge organized by the 2019fourth Conference on Machine Translation (WMT19), out of the eleven tasks itundertook. Overall, there are nineteen machine translation categories in WMTthis year.
MSRA achieved first place in machine translation tasksfor Chinese-English, English-Finnish, English-German, English-Lithuanian, French-German,German-English, German-French and Russian-English. Three other tasks wereplaced second in their respective categories, which included English-Kazakh,Finnish-English and Lithuanian-English.
As one of the leading machine translation competitionglobally, WMT is a platform for leading researchers to demonstrate their solutions,as well as to understand the continuous evolvement of machine translationtechnology. Now in its 14th year, more than 50 teams globally fromtechnology companies, leading academic institutions and universities participatedin a bid to demonstrate their machine translation capabilities.
The organizers aimed to evaluate current machinetranslation techniques for the languages other than English, as well as toexamine the challenges between European languages, including low resource andmorphologically rich languages.
Improvements toMulti-dimensional Algorithms for Better Machine Translation Outcomes
"This year,the MSRA team applied innovative algorithms to its system, which significantly improvedthe quality of the machine translation results. These algorithms were used toimprove the platform's learning mechanism, pre-training, network architectureoptimization, data enhancement and other processes required so that the systemcan perform better," explains Tie-Yan Liu, Assistant Managing Directorof MSRA.
The innovative algorithms leveraged this year include:
- MADL：Multi-agent dual learning
- MASS：Masked sequence to sequence pre-training
- NAO：Automatic neural architecture optimization
- SCA：Soft contextual data augmentation
The achievement follows the 2018 breakthrough whereby researchers in MSRA and Microsoft Research U.S. labs reached human parity on a commonlyused test set of news stories, callednewstest2017, which wasdeveloped by a group of industry and academic partners and released at WMT17.The system is able to translate sentence of news articles from Chinese toEnglish with the same quality and accuracy as a person.
"The realm of machine translation will continue toevolve with better algorithms, data set and technology. However, much of our research today is really inspired byhow we humans do things. Language is complex and nuanced, as people can usedifferent words to express the exact same concept. Hence, developingmulti-dimensional algorithms is important in evolving machine translationsystems so that they can deliver better outcomes," said Liu. "Ourachievement at WMT19 serves to the further development of the field, whereby wehope that machine translation can become better in the years to come."
For example, Microsoft Translator,a multilingual machine translation cloud service, has integrated some of theprevious solutions developed by Microsoft Research teams globally to enhancethe accuracy of the tool. Now, the research teams plan to integrate the newalgorithms used for this year's WMT challenge to improve its offering.
Microsoft(Nasdaq "MSFT" @microsoft) enables digital transformation for the era of anintelligent cloud and an intelligent edge. Its mission is to empower everyperson and every organization on the planet to achieve more.