[CL+NLP Lunch] Tuesday: Chris Dyer, morphology in machine translation

[Nathan Schneider]

Everyone,

On Tuesday, Chris Dyer of LTI will speak to the CL+NLP Lunch. Details
are below. (He plans to finish by 1:00.)

Ben & Nathan


Tuesday, Jan. 18 @ noon
GHC 6115

Chris Dyer
Postdoctoral Fellow, LTI

Inflectional Morphology in Probabilistic Translation Models

Abstract:
In conventional translation models, words that differ from each other
in any way are modeled independently of each other. From a modeling
perspective, this is unsatisfying since closely related morphological
forms of an underlying stem are likely to share many characteristics
that are important for translation. And, more practically, this
independence assumption means data sparsity is a significant issue in
translation between morphologically complex languages.

I compare two new probabilistic translation models that relax this
"lexical independence assumption" and share statistics across
morphologically related word forms. The first model is generative,
based on hierarchical Pitman-Yor processes, in which the translation
distributions for different inflection variants of a stem share a
common base distribution. The second model is based on Markov random
fields and uses morphological features to share information across
related forms.

Lunch will be provided.