AI Seminar
    -------------------------------
    Tuesday, September 23rd, 2003
    4:00 pm - 5:30 pm
    175 ATL (Large Conference Room)

    "Exploiting sentence paraphrasing for machine translation OR machine
    translation using alternative phrasings of the same sentence"

    Dragomir Radev
    Department of Electrical Engineering and Computer Science
    University of Michigan


----------------------------------
Recent work in statistical machine translation has moved from simple
word-based alignment models to higher-order models involving syntax
(e.g, alignment templates [Och,Tillman&Ney99; Och02] or tree-to-string
models [Yamada&Knight02]. Such models incorporate syntactic knowledge
directly into the translation process.  A different technique,
syntax-based reranking, was recently proposed for improving quality of
machine translation. Reranking (e.g., [Radev,Prager,Samn00;
Collins00]) is based on a discriminative model (e.g., using maximum
entropy) over a large list of candidates.

A recent 13-person project, led by Franz Josef Och of ISI, looked at
syntax-based reranking to incorporate a large number of (mainly)
syntactic features in the reranking model for a large-scale
Chinese-to-English machine translation system.

I will present an overview of the project and then turn to my
individual contributions - namely two sets of features based on (1)
alternative translation paraphrases and (2) lopsided (flipped) lexical
dependencies. I will present evaluation results using the official
NIST evaluation corpus that show the viability of these features and
of the approach in whole.