[Intelligence Seminar] May 15: Dan Klein, NSH 1305, 2:00 - "Latent-Variable Models for Natural Language Processing"

[Noah A Smith]

Joint Intelligence/LTI Seminar

May 15, 2009 (note special time and place)
2:00 pm
NSH 1305
For meetings, contact Noah Smith (nasmith at cs.cmu.edu).

Latent-Variable Models for Natural Language Processing
Dan Klein
Computer Science Division, University of California, Berkeley

Abstract:

Language is complex, but our labeled data sets generally aren't.  For
example, treebanks specify coarse categories like noun phrases, but
they say nothing about richer phenomena like agreement, case,
definiteness, and so on.  One solution is to use latent-variable
methods to learn these underlying complexities automatically.  In this
talk, I will present several latent-variable models for natural
language processing which take such an approach.

In the domain of syntactic parsing, I will describe a state-splitting
approach which begins with an X-bar grammar and learns to iteratively
refine grammar symbols.  For example, noun phrases are split into
subjects and objects, singular and plural, and so on.  This splitting
process in turn admits an efficient coarse-to-fine inference scheme,
which reduces parsing times by orders of magnitude.  Our method
currently produces the best parsing accuracies in a variety of
languages, in a fully language-general fashion.  The same techniques
can also be applied to acoustic modeling, where they induce latent
phonological patterns.

In the domain of machine translation, we must often analyze sentences
and their translations at the same time.  In principle, analyzing two
languages should be easier than analyzing one: it is well known that
two predictors can work better when they must agree.  However
``agreement'' across languages is itself a complex, parameterized
relation.  I show that, for both parsing and entity recognition,
bilingual models can be built from monolingual ones using
latent-variable methods -- here, the latent variables are bilingual
correspondences.  The resulting bilingual models are substantially
better than their decoupled monolingual versions, giving both error
rate reductions in labeling tasks and BLEU score increases in machine
translation.


Bio:

Dan Klein is an assistant professor of computer science at the
University of California, Berkeley (PhD Stanford, MSt Oxford, BA
Cornell).  His research focuses on statistical natural language
processing, including unsupervised learning methods, syntactic
parsing, information extraction, and machine translation.  Academic
honors include a Marshall Fellowship, a Microsoft New Faculty
Fellowship, the ACM Grace Murray Hopper award, and best paper awards
at the ACL, NAACL, and EMNLP conferences.