dcsimg

 SDL Research 
 Driving Translation Innovation 

About

SDL is leading innovation in language technology to bring down global communication barriers to support international business growth and cross language interaction.  SDL Research is revolutionizing what is possible through ongoing advancements in statistical prediction on big data and statistical learning from machine/human synergy.

Located across two innovation centers – Los Angeles, California and Cambridge, England – SDL Research is made up of world-renowned researchers that contribute to an exciting research agenda based on the current needs of SDL’s customers and what is needed to drive the future of global communication.

Innovation

The solutions to the hard science problems being tackled by SDL Research regularly make their way into SDL products to support business requirements. A few of these innovations are highlighted below.

Statistical machine translation

Several researchers at SDL Research were part of the team at SDL Language Weaver, the first company to commercialize a statistical approach to machine translation (MT) that improves the quality and applicability of MT.  Statistical Machine Translation (SMT) is now the leading approach to machine translation.  With this approach, the computer software learns how to translate new content by learning statistical models from existing, human translated parallel texts.  This approach enables the production of more fluent translations, faster creation of new language pairs, and easier customization to improve quality in specific domains. The statistical machine translation algorithms developed by the team power the SDL Enterprise Translation Server, SDL Government Platform, and SDL BeGlobal products; and FreeTranslations.com.

Trusted Translations

An ongoing challenge for users of machine translation is knowing whether the quality is “good enough.” SDL Research took on this challenge and developed a confidence estimation algorithm, called TrustScore. This algorithm employs automatic learning techniques, trained over thousands of previously translated examples and learns correlations between automatic translations and their perceived human quality/utility. The algorithm is then deployed to infer utility scores for automatic translations of new, previously unseen texts. This innovation was added to SDL’s machine translation platform, SDL BeGlobal, and delivers a quality score between 1 and 5 for each translation so that users know if the translation is good enough to publish or if it requires additional review.

Machine translation “training”

One of the powerful characteristics of statistical MT is the ability to easily “train” it for a particular domain. The SDL Research team has packaged the experience gained while training tens of thousands of engines into SDL BeGlobal Trainer, an extension of SDL BeGlobal. BeGlobal Trainer enables users to customize language pairs on their own, without involving SDL. 

Automated Translation