SCS Faculty Candidate Talk

  • Gates&Hillman Centers
  • ASA Conference Room 6115
  • GRAHAM NEUBIG
  • Assistant Professor
  • Graduate School of Information Science
  • Nara Institute of Science and Technology (NAIST)
Seminars

What can Neural Machine Translation Learn from Symbolic Methods?

In this talk, I will compare and contrast two major paradigms for statistical machine translation (MT), and explain our ongoing work on methods to unify the two. The first and more traditional paradigm, using symbolic methods, work by dividing sentences up into small chunks, translating these chunks, and re-combining the chunks into a full sentence translation. The second paradigm, using neural methods, generates the target words one-by-one according to a probabilistic model that considers the source sentence and the previous words in the target.

First, I will overview the symbolic methods, specifically focusing on our work on syntax-based MT that has allowed us to achieve state-of-the-art results on language pairs with very different word orders. Given this background, I will describe neural MT, contrasting it to the more traditional symbolic methods, and showing results demonstrating the relatively new neural MT paradigm's enormous potential, but also pointing out some quirks that seriously impact its usability in real MT applications. Finally, I will present a new technique that we have recently developed to incorporate information from symbolic models into neural models, and describe ongoing work to use this and similar techniques to create neural MT systems that are as flexible and reliable as their symbolic counterparts.

Graham Neubig am currently an assistant professor at the Nara Institute of Science and Technology (NAIST) Graduate School of Information Science, affiliated with the Augmented Human Communication Laboratory.  I completed my PhD at Kyoto University in 2013 with Prof. Tatsuya Kawahara. 

My research focuses on handling human languages (like English or Japanese) with computers, so called natural language processing.  In particular, I am interested in machine learning approaches that are both linguistically motivated and tailored to applications such as machine translation and speech recognition.  I like developing and have created open-source software.  I also have had a chance to give tutorials and talks on subjects such as machine translation, neural networks, and language modeling.

Faculty Host: Alan Black

For More Information, Please Contact: 
Keywords: