MT Marathon 2015 Keynote Talks
Monday through Friday, MT Marathon includes keynote talks.
For slides and videorecordings (where available), see the programme.
Confirmed Speakers
Tomáš Fulajtár (Moravia IT)
Real-World Application of an Machine Translation Workflow
Exponential growth in content and shorter publishing cycles in business environments require faster and high-quality translations. Moravia works with the top brands in the world and is faced daily with the challenge of continuous localization of content. Machine Translation plays a key role when speed is of the essence and quality is not negotiable. Customized workflows, bespoke tools, and documented processes help in creating an environment that increases productivity over time and addresses the unique business needs of each of our clients. Tomas will share with the MT community some of the learnings gained from the real-world application of Machine Translation to solve business challenges.
François Yvon (LIMSI, University Paris Sud)
From n-gram Translation Models to large-scale, discriminatively trained conditional, Translation Models
Over the past decade or so, LIMSI has been developing its own machine translation technology, generally referred to as "n-gram based" machine translation - owing to the use of a n-gram bilingual model as its main translation model (note that this approach was initially developped by other groups, notably at UPC Barcelona). In this talk, I will try to recap these developments, highlighting the main peculiarities of the n-gram based approach as compared to more conventional SMT architectures. I will then focus on recent attempts to move from n-gram models, which are typically trained with likelihood-based criteria, to large-scale, discriminatively trained translation models. Two possible lines of developments have been explored: using neural nets, on the one hand; using conditional random fields, on the other hand. Analyzing where and why these approaches have succeeded, or failed, I will try to outline further prospects and chart the way forward for n-gram-based SMT.
Keith Stevens (Google)
Neural Network Models in Google Translate
Google Translate supports over 90 languages with world class quality for many of them. With normal Phrase Based models, we're reaching a limit of what's possible, especially for languages vastly different from English. This leaves many language pairs with disappointing quality until now. How can we go beyond the limitations of phrase based models? With Neural Networks. These enable richer features and longer dependencies. Some models support standard phrase based models and some replace them entirely. I'll cover a variety of Neural Network Machine Translation models we're researching at Google with some preliminary results.
Hinrich Schütze (Ludwig-Maximilians-Universität München)
Text Representations for NLP and MT
For statistical NLP in general and for statistical MT in particular, the representation of the input text is key: many important regularities of translation can only be learned if the text is represented in a form that is conducive to generalization and reduces sparseness. I will review recent work in my group in two areas that can contribute to text representations that support learning effective statistical models: morphological analysis and deep learning embeddings.
Walther von Hahn, Wolfgang Menzel (Uni Hamburg)
as part of the shared session with the Prague-Hamburg workshop on Friday