Predicting the quality of an MT engine without existing target reference is one of the tricky part in MT technology. It plays an essential role in making MT usable in real life scenarios. Perspective by Gábor Bessenyei (CEO of MorphoLogic Localisation Ltd.).
The music-loving Baltic countries are a multilingual hotspot in Europe, with the majority of citizens speaking (and singing) three languages on a daily basis. At the same time, the melodious Baltic languages are famously complex and morphologically rich, containing lots of ambiguity and intricate word agreements. Taken together, these factors make the region a prime spot for driving innovation in language technologies. Tilde, a language technology company specializing in custom MT and terminology services, has leveraged its extensive linguistic experience in the Baltic region to create custom MT systems for a wide variety of languages and domains, helping EU and global companies to boost translation productivity and make their applications multilingual. Tilde recently embarked on the challenging task of building a large-scale MT service for the Latvian government, Hugo.lv. This service was adapted to create a communication tool for the 2015 EU Presidency. The presentation will introduce the audience to languages and MT in the Baltic region and highlight these two case studies, which showcased the crucial role of language technology in enabling multilingual communication in the digital age.
Tony O’Dowd (KantanMT). KantanMT enables its community to generate meaningful business intelligence that helps them identify the scope of their customised machine translation projects. More importantly, it helps them schedule and scale those projects to achieve maximum translation productivity and a positive ROI.
In all of our translation production activities we are producing data, lots of data. We are not talking now about the actual translations that are stored as translation memory data. These translation memory data have proven to be very valuable over the years and recently again as training data for Machine Translation engines. But in this session we are talking about the other data: data about the translation process. How much time was spent on different tasks, for different languages, content types, per project? What was the quality score for the translator, for the vendor? What was the user feedback on this machine translated support article? How is our MT engine performing? And has it improved since last year, since we have added 13 million more words in the training set? Some of the buyers and providers of translation are further ahead with the use of all these translation management data than others. The TAUS Dynamic Quality Framework (DQF) tracks translation management data through plug-ins that are already available for various translation tools and platforms. The vision is becoming very clear: the translation industry can have its own “Big Data”. In the past couple of months TAUS enterprise members have contributed their wishes and requirements for an industry benchmarking platform for translation quality and productivity. In this session several TAUS members will share and discuss their plans for using DQF and the Quality Dashboard. What data would you like to track?
Session host: Daniel Goldschmidt (Microsoft)
Presenters and panelists are: Annya Sedakova-Bertram (EMC), Fred Tuinstra (Lionbridge), Achim Ruopp (TAUS)
In this session, with clear focus on Machine Translation (MT) quality, we will discuss different ways to improve MT engines. Which engine do you use and how do you measure improvement? What are the right metrics to evaluate MT quality for the specific content types? How do you interpret and act on the evaluation results? It's fine when errors are labeled and analyzed, but how can that help improve your engine? Are there best practices available? And how about Neural MT? Should we measure that differently? After some use cases shared by the speakers, these questions will be addressed in the break-out session.
Predicting the quality of an MT engine without existing target reference is one of the tricky part in MT technology. It plays an essential role in making MT usable in real life scenarios. Perspective by Gábor Bessenyei (CEO of MorphoLogic Localisation Ltd.).
The music-loving Baltic countries are a multilingual hotspot in Europe, with the majority of citizens speaking (and singing) three languages on a daily basis. At the same time, the melodious Baltic languages are famously complex and morphologically rich, containing lots of ambiguity and intricate word agreements. Taken together, these factors make the region a prime spot for driving innovation in language technologies. Tilde, a language technology company specializing in custom MT and terminology services, has leveraged its extensive linguistic experience in the Baltic region to create custom MT systems for a wide variety of languages and domains, helping EU and global companies to boost translation productivity and make their applications multilingual. Tilde recently embarked on the challenging task of building a large-scale MT service for the Latvian government, Hugo.lv. This service was adapted to create a communication tool for the 2015 EU Presidency. The presentation will introduce the audience to languages and MT in the Baltic region and highlight these two case studies, which showcased the crucial role of language technology in enabling multilingual communication in the digital age.
Tony O’Dowd (KantanMT). KantanMT enables its community to generate meaningful business intelligence that helps them identify the scope of their customised machine translation projects. More importantly, it helps them schedule and scale those projects to achieve maximum translation productivity and a positive ROI.
In all of our translation production activities we are producing data, lots of data. We are not talking now about the actual translations that are stored as translation memory data. These translation memory data have proven to be very valuable over the years and recently again as training data for Machine Translation engines. But in this session we are talking about the other data: data about the translation process. How much time was spent on different tasks, for different languages, content types, per project? What was the quality score for the translator, for the vendor? What was the user feedback on this machine translated support article? How is our MT engine performing? And has it improved since last year, since we have added 13 million more words in the training set? Some of the buyers and providers of translation are further ahead with the use of all these translation management data than others. The TAUS Dynamic Quality Framework (DQF) tracks translation management data through plug-ins that are already available for various translation tools and platforms. The vision is becoming very clear: the translation industry can have its own “Big Data”. In the past couple of months TAUS enterprise members have contributed their wishes and requirements for an industry benchmarking platform for translation quality and productivity. In this session several TAUS members will share and discuss their plans for using DQF and the Quality Dashboard. What data would you like to track?
Session host: Daniel Goldschmidt (Microsoft)
Presenters and panelists are: Annya Sedakova-Bertram (EMC), Fred Tuinstra (Lionbridge), Achim Ruopp (TAUS)
In this session, with clear focus on Machine Translation (MT) quality, we will discuss different ways to improve MT engines. Which engine do you use and how do you measure improvement? What are the right metrics to evaluate MT quality for the specific content types? How do you interpret and act on the evaluation results? It's fine when errors are labeled and analyzed, but how can that help improve your engine? Are there best practices available? And how about Neural MT? Should we measure that differently? After some use cases shared by the speakers, these questions will be addressed in the break-out session.