[CL+NLP Lunch] CL+NLP Lunch Tuesday Dec 4 @ noon

[Dani Yogatama]

Hi all,

Dirk Hovy's talk has been rescheduled for Tuesday December 4 at
noon. Details are included below.
Lunch will be provided.


Thanks,
Dani

P.S.: You may also be interested in Nathan Schneider's thesis proposal
tomorrow at 5pm (http://calendar.cs.cmu.edu/scsEvents/demo/8238.html).

===================================================
*CL+NLP Lunch* (http://www.cs.cmu.edu/~nlp-lunch/)
*Speaker*: Dirk Hovy, University of Southern California
*Date*: Tuesday, December 4, 2012
*Time*: 12:00 noon
*Venue*: GHC 6115

*Title*:
Learning Semantic Types and Relations from Syntactic Context

*Abstract*:
Natural Language Processing (NLP) is moving towards incorporating more
semantics into its applications, such as Machine Translation (MT) or
Question Answering (QA). Most semantic frameworks depend on predefined
"external" resources, such as knowledge bases or ontologies. This requires
a lot of manual effort, but even then it is impossible to build a complete
representation of the world. Instead, we would like to learn a sufficient
representation directly from data.

One way to encode meaning is through syntactic and semantic relations
between predicates and their arguments. Relations are thus at the core of
meaning and information. Recently, several approaches have collected large
corpora of syntactically related word chains (e.g., subject, verb, object).
However, these chains are extracted at the lexical level and do not
generalize well, so their use for semantic interpretation is limited. If we
could use these lexical chains as inputs to generalize beyond the word
level, we would be able to learn semantic relations and make use of these
existing resources.

In this talk, I will present a method to learn semantic types and relations
from raw text.
I construct unsupervised models on syntactic dependency arcs, using
potential types as latent variables. The resulting models allow for quick
domain adaptation and unknown relations, and avoid data sparsity caused by
intervening words.
I show improvements over state-of-the-art systems as well as novel
approaches to fully exploit the structure contained in the data. My method
builds on existing triple stores and does not require any external
knowledge bases, manual annotation, or pre-defined predicates or arguments.

*Bio*:
Dirk Hovy is a PhD candidate in Natural Language Processing at the
University of Southern California (USC), and holds a Masters in
linguistics. He is interested in data-driven models of meaning and
understanding and has published on unsupervised learning, information
extraction, word-sense disambiguation, and temporal relation classification
(see <http://www.dirkhovy.com/portfolio/papers/index.php>).
His thesis work is concerned with how computers can learn semantic types
and relations from raw text, without recurrence to external resources.
His other interests include baking, cooking, crossfit, and medieval art and
literature.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/nlp-lunch/attachments/20121127/dbf73afd/attachment.html>