Tuesday, April 26, 2011
9:00 AM
Computer Science Bldg. Room 601
Ph.D. Thesis Proposal
Elif Eyigoz
University of Rochester
Explicit Morphology Modeling for SMT
All current state of the art approaches to SMT rely on an automatically word-aligned corpus.
However, the vast majority of human languages exhibit some degree of morphology. Since
morphology has not been addressed explicitly in the current SMT models, machine translation
of morphologically rich languages has proven difficult. Researchers have tried ways to tweak
SMT systems to perform better with morphologically richer languages by manipulating the
content and the form of what should be a word, i.e. the smallest unit of translation.
The IBM models have become the principal method for inducing word-alignments for other
translation models. We present a new alignment model that jointly induces word and morpheme
alignments, using an EM algorithm. The model is a generalization of the IBM models. That is,
it represents a word as a one-level tree, whereas the IBM models represent a word as a single
node. In future work, we plan to model a word as a tree of arbitrary height, and utilize this
information in a syntax-based approach to SMT.