Monday, January 03, 2011
1:00 PM
Computer Studies Bldg. Room 703
Shaojun Zhao
University of Rochester
On Word Alignment Models for Statistical Machine Translation
Machine translation remains the holy grail of computational
linguistics. All statistical ma- chine translation systems are built
upon the idea of word alignment. While the field of word alignment has
had tremendous progress in the last two decades, it is still in great
need of speed and quality improvement.
We designed a fertility hidden Markov model for word alignment, which
is dramatically faster than the most widely used IBM Model 4. In fact,
our model is even faster and has lower alignment error rate (AER) than
the hidden Markov model. An experiment on Chinese-English translation
shows that our word alignment model leads to better translation
results than IBM Model 4, based on the BLEU metric.
We also designed algorithms that mine massive and high quality
bilingual texts for a variety of language pairs from the web using
word alignment. The resulting data improved a state-of- the-art
machine translation system.