Roland Kuhn
Mon 20 Apr 2015, 11:00 - 12:00
Informatics Forum (IF-4.31/4.33)

If you have a question about this talk, please contact: Diana Dalla Costa (ddallac)


In the last two years, models that exploit richer information from the source sentence have significantly improved the quality of outputs from statistical machine translation (SMT) systems. At the National Research Council of Canada (NRC), we have recently implemented two such models – one homegrown, one a reimplementation of someone else’s work.

These two models are
1. Coarse biLMs – a version of the Niehues et al bilingual LM (2011) in which components are automatically clustered in several different ways.
2. NNJM – the Neural Network Joint Model of Devlin et al (2014).

We obtained large BLEU improvements over a state-of-the-art phrase-based baseline from both of these techniques for Arabic > English and Chinese > English tasks within the DARPA BOLT project (for coarse biLMs, we also carried out English > French and French > English experiments, and saw good improvements for both). For reasons we do not fully understand, our group’s reimplementation of NNJM seems to have been more successful than similar efforts elsewhere – that is, we obtain improvements of the same magnitude as those reported in the Devlin et al paper (though we have not yet implemented all of the techniques in that paper).

The presentation will discuss coarse biLMs and NNJMs, while touching on related approaches such as Factored Markov Translation (Feng, Cohn, and Du 2014) and the University of Edinburgh’s Operation Sequence Model (which when incorporating generalized word representations, has a philosophical similarity to coarse biLMs). Finally, we will discuss preliminary results for experiments in which coarse biLMs and NNJMs were combined in the same system, in order to determine to what extent the information they exploit is overlapping or complementary.

Since Jan. 2014, Roland Kuhn has been Team Leader for the Multilingual Text Processing group at the National Research Council of Canada (NRC). He has been Co-Leader of the Portage machine translation project at NRC since July 2004. After studying mathematical biology at the University of Toronto and the University of Chicago, Dr. Kuhn developed an interest in natural language. In 1993, he received his Ph.D. in Computer Science from McGill University. He worked at the Centre de recherche informatique de Montréal (CRIM) as a researcher and a senior researcher between 1992 and 1996, then held research and development positions with Panasonic Speech Technology Laboratory in Santa Barbara, California from October 1996 to June 2004. He joined the National Research Council of Canada (NRC) in 2004. In the course of his research career, he has studied a diverse set of problems in natural language processing, including automatic speech recognition and machine translation.