Thursday, May 05, 2011
2:00 PM
Computer Studies Bldg. Room 703
Ph.D. Thesis Proposal
Licheng Fang
University of Rochester
Accurate and Efficient Synchronous Context-Free Grammars
Synchronous context-free grammar (SCFG) is an expressive yet expensive model for
machine translation. In this thesis proposal, we suggest several areas of improvements
in both learning SCFGs and decoding with SCFGs. First, we propose a grammar binarization
method that makes decoding with SCFGs more efficient by taking into account
terminal information. Second, we propose ways to learn a compact and accurate SCFG
from word-aligned bilingual data. Lastly, we propose to decode SCFGs with a new
dynamic programming scheme that generates English translations from left to right, which
reduces language model integration complexity and may result in better running time in practice.