• Phone +(306) 222-3339
  • E-mail Address email@yourcompany.com

AMTA 2016 | Keynotes & Panel

Saturday Morning:

Rico Sennrich, University of Edinburgh
Neural Machine Translation: Breaking the Performance Plateau [Slides here]

Neural machine translation has seen substantial improvements in the last  two years, and is quickly establishing itself as the dominant MT technology. I will present recent advances that have contributed to the rapid progress in the area. I will then analyze the strengths of neural machine translation, and how it has been able to surpass phrase-based methods in aspects such as fluency, and discuss challenges that the MT community will need to address in the future.

Rico Sennrich is a research associate at the Institute for Language, Cognition and Computation, University of Edinburgh, where he has worked since 2013. He received his PhD in Computational Linguistics from the University of Zurich in 2013. He works on data-driven natural language processing, in particular machine translation, syntax and morphology. At the WMT 2016 shared news translation task, his neural machine translation submission was ranked (tied) best for 7 translation directions.
Mike Dillinger, LinkedIn
MT Escaped from the Lab. Now What? [Slides here]

The MT landscape has changed radically in the last 10 years.  For a very long time the key bottleneck for more widespread use of MT was ill-informed users; now MT has finally become the standard solution for several large-scale commercial use cases:  support content, user feedback, cross-language search, and others.  Users are clamoring for much more and much better MT.  The tables have turned: these days the researchers are the main bottleneck. Today we see huge, ferocious MT systems that have escaped from labs around the world and made their way into innocent, unsuspecting organizations. These systems savagely translate everything in their paths, sowing confusion among their handlers and creating panic in managers everywhere. Users can’t tame or control them; they can’t predict when they’ll work well or create a mess; they can’t see what’s going wrong or  fix it. Today, from the users’ standpoint MT is a plug-and-pray technology. The underlying cause of this situation is the fact that today MT researchers and MT users define the problems that they want to solve in radically different ways: there is a dramatic mismatch between what commercial MT users need and what MT researchers offer.  This talk identifies and describes these differences so that together we can address a wider range of practical use cases and finally see MT in use everywhere.

Mike Dillinger, PhD manages Taxonomies and Machine Translation at LinkedIn. He is a two-time Past President of AMTA, has been on the Board for about 15 years, and serves on the advisory board of several startups. He led the first deployments of MT at both eBay and LinkedIn, and he worked as an independent consultant helping clients like Apple, HP, Cisco, and others to use MT effectively. Dr. Dillinger started in MT by studying with early pioneers Paul Garvin (chief linguist on the Georgetown MT project) and David Hays (director of the MT project at the Rand Corporation and a member of the ALPAC Commission), then worked with Hiroshi Uchida (creator of the Atlas System at Fujitsu) and Bud Scott (founder of Logos Corporation). He worked with interlingual MT in the multinational UNL Project, with commercial rule-based MT at eBay, Spoken Translation, GlobalWords Technologies, and Logos Corporation, and with statistical MT at eBay and LinkedIn. He published a wide range of articles, contributed to the emerging standards OLIF and UNL, and was awarded two patents for translation technology. Dr. Dillinger has taught at more than a dozen universities in several countries and has been a visiting researcher on four continents.
Sunday Morning:
Spence Green, Lilt
Interactive Machine Translation: From Research to Practice [Slides here]

In the computer-aided translation setting, there are two classic problems with batch-trained, full-sentence machine translation. First, the translation system repeats mistakes, even after those mistakes are corrected by the user. Second, the system cannot easily refine its suggestions given partial user feedback, restricting interface choices to variants of basic post-editing, which users have not historically preferred. Interactive machine translation aims to solve these problem with adaptive translation models and human-centric, predictive interfaces. This talk presents three generations of interactive translation systems, two of which were built at Stanford and the third at Lilt. I will focus on recent research on prefix-based tuning and decoding, and also present preliminary results on neural models for this task. Finally, I will describe recent interface choices in the framework of mixed-initiative design.

Spence Green is a co-founder of Lilt. He received a PhD in computer science from Stanford University in 2014. He also holds a BS in computer engineering from the University of Virginia.
Douglas A. Jones, MIT Lincoln Laboratory
DARPA Human Language Technology Programs  [Slides here]

Presentation of Boyan Onyshkevych’s Human Language Technology Programs at DARPA: DEFT, LORELEI, and a Potential New Program Space.  Details for DEFT (Deep Exploration and Filtering of Text) include English, Spanish and Chinese examples, Research Areas, NIST Open Evaluation Task Alignment, and DEFT Program Participants.  Details for LORELEI (Low Resource Languages for Emergent Incidents) include: Program Motivation; Potential HLT Use Case; Program Goal; Example Use Case: Mission Planning; Example Use Case: Situation Awareness; Hypothetical Mission Requirements and Answers; Program Concept; Performer Teams; and Performer Main Focus Areas.  Details for BOLT (Broad Operational Language Translation) include: Available BOLT Software and New BOLT Transitions.

Doug Jones is a member of the technical staff in the Human Language Technology Group at MIT Lincoln Laboratory. His background includes B.A. and M.A. degrees in linguistics from Stanford University specializing in computational phonology, and a Ph.D. in linguistics from the Massachusetts Institute of Technology specializing in Hindi syntax. He completed postdoctoral work on computational theories of verb structure at MIT’s Artificial Intelligence Laboratory, and the University of Maryland’s Institute for Advanced Computer Studies (UMIACS). Dr. Jones has held research positions in the U.S. government with the Department of Defense where he specialized in machine translation for world minority languages, and at National Institute of Standards and Technology where he helped launch a Chinese-English cross language information retrieval study. The main focus of his research is leveraging the inherent structure of linguistic patterns for the design of large-scale human language processing systems. Since 2002, a major focus has been to adapt military standards of foreign language testing for machine translation evaluation. He is working to develop a common measure both for human language learners and for machine translation technology, the purpose being to influence technology in ways that best enable people to accomplish foreign language tasks. He has published numerous papers in the field of computational linguistics.  He was part of the U.S. Southern Command (USSOUTHCOM) and Joint Task Force-HAITI team that responded to the January 2010 earthquake in Haiti. His publications and other information are available at: http://www.mit.edu/~dajones.
Monday Morning:
Panel on “MT Commercialization: Past, Present and Future”
Moderator:
Daniel Marcu, Director of Strategic Initiatives, ISI/USC, and Founder, FairTradeTranslation.com
Panelists:
Macduff Hughes, Engineering Director, Google Translate
Valery Jacot, Software Development Manager, Localization Platform, Autodesk
Dragos Munteanu, Director of Research and Development MT, SDL plc.

Chris Wendt, Group Program Manager, Machine Translation, Microsoft

You, Contribute to the panel by responding to this questionnaire.

Summary:

In the early 2000s, the dominant company in the field, Systran, was generating approximately 10M Euros in Machine Translation revenues while translating relatively small amounts of text; and a slew of MT-focused startups have crashed and burned after spending millions of dollars of investors’ money.  By 2016, Google, Microsoft, and Facebook have created services that cumulatively translate  more than 150 Billion words per day; and the most successful MT-focused startups have been acquired by larger businesses. A panel of experienced MT commercialization experts will review lessons learned during the last 15 years while bringing MT to market and/or using MT to create value inside their own organizations. They will review the commercial segments in which MT is good enough to generate value; barriers to adoption; and the current trends that keep them awake at night. They will also provide their perspective on the future of this technology and the likely evolution of the market: how many MT providers will we have 10 years from now?

 

Moderator:

  • Daniel Marcu is a Director of Strategic Initiatives at ISI/USC and a Founder at FairTradeTranslation.com. He enjoys creating and growing commercial organizations; developing advanced R&D concepts, and transitioning them into novel commercial software and services. In the past, he has co-founded and built Language Weaver Inc. and has created and contributed to software that was used by more than 100 million people worldwide. He is a Fellow of the Association for Computational Linguistics for significant contributions to discourse parsing, summarization, and machine translation and for kick starting the statistical machine translation industry.

Panelists:

  • Macduff Hughes has led the Google Translate team as Engineering Director since 2012. He has worked at Google since 2007, having previously led the Google Voice and Google Accounts teams. He has a bachelor’s degree from Stanford University and did graduate studies at the University of Trier and Columbia University.
  • Valery Jacot is a Software Development Manager at Autodesk, where he manages the development and the operational effort for the Autodesk Translation Platform. He graduated from the West Switzerland University of Applied Science in Software Development and worked as Software Developer and Software Architect in several businesses such as private banks, insurances companies, and, more recently, in the Autodesk Localization/internalization area. His goal is to improve translation experience within the Autodesk business.
  • Dragos Munteanu is Director of Research and Development for Machine Translation at SDL. Dragos is currently managing the end-to-end development life cycle of SDL’s Statistical Machine Translation Products, with a focus on continuous translation quality improvement via innovation in algorithms and advancement in scalability. Dragos has extensive knowledge in Statistical Machine Translation, Machine Learning, and Natural Language Processing. He has 10+ years of experience in the translation industry with significant contributions as both scientist and product manager. Dragos has a Ph.D. in Computer Science from the University of Southern California, and an MBA from the University of California Los Angeles.
  • Chris Wendt is a Group Program Manager, Machine Translation, at Microsoft. He graduated as Diplom-Informatiker from the University of Hamburg, Germany, and subsequently spent a decade on software internationalization for a multitude of Microsoft products, including Windows, Internet Explorer, MSN and Windows Live – bringing these products to market with equal functionality worldwide. Since 2005 he is leading the program management team for Microsoft’s Machine Translation development, responsible for Microsoft Translator services, including Bing Translator and Skype Translator, connecting Microsoft’s research activities with its practical use in services and applications. He is based at Microsoft headquarters in Redmond, Washington.