The CLIP Colloquium Series presents...

Scalable Discriminative Learning for Powerful Translation Models

I. Dan Melamed (NYU)
September 21, 2006, 1:30pm, AVW 3258 NOTE SPECIAL TIME AND PLACE

The translational equivalence relations in a surprisingly high fraction of ordinary bitexts cannot be effectively explained by commonly used translation models. Models of translational equivalence with more expressive power are necessary. However, currently popular machine learning techniques do not scale up well for these more powerful models, which limits these models' practical utility. I shall present a new purely discriminative learning method for structured prediction problems, including parsing and translation. This method scales up to millions of features over large training sets, such as those used for statistical MT. Experiments have shown that even context-sensitive models can be effectively trained using this method. If there is interest, I will also give an overview of the GenPar toolkit, which made all of this work possible.

RELEVANT PAPERS:

(all but the first are online at http://www.cs.nyu.edu/~melamed/pubs.html; GenPar is available from http://nlp.cs.nyu.edu/GenPar.)

• Joseph Turian, Benjamin Wellington, and I. Dan Melamed (to appear in 2006). Scalable Discriminative Learning for Natural Language Parsing and Translation, in NIPS.
• Benjamin Wellington, Joseph Turian, Chris Pike, and I. Dan Melamed (2006). Scalable Purely-Discriminative Training for Word and Tree Transducers, in AMTA.
• Benjamin Wellington, Sonjia Waxmonsky, and I. Dan Melamed (2006). Empirical Lower Bounds on the Complexity of Translational Equivalence, in ACL.
• Joseph Turian and I. Dan Melamed (2006). Advances in Discriminative Parsing. in ACL.
• A. Burbank, M. Carpuat, S. Clark, M. Dreyer, P. Fox, D. Groves, K. Hall, M. Hearne, I. D. Melamed, Y. Shen, A. Way, B. Wellington, and D. Wu (2005) Final Report of the 2005 Language Engineering Workshop on Statistical Machine Translation by Parsing
• I. Dan Melamed (2004). Statistical Machine Translation by Parsing, in ACL.
• I. Dan Melamed, Giorgio Satta, and Benjamin Wellington (2004). Generalized Multitext Grammars, in ACL.

About the Speaker

Dan Melamed received his Ph.D. from the University of Pennsylvania in 1998. After spending three years in industry, he is now an assistant professor at NYU.

This talk is part of the CLIP Colloquium Series, organized by Jimmy Lin (jimmylin -at- umd .dot. edu). For the complete schedule, please visit http://www.umiacs.umd.edu/research/CLIP/colloq/.