[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[eu_members at aclweb dot org] Colorado: NAACL HLT 2009 -- Workshop DEADLINE EXTENSION: Integer Linear Programming for NLP

Subject: Extended Deadline (15 March): Integer Linear Programming for NLP Workshop at NAACL-HLT 2009


Workshop on Integer Linear Programming for NLP at NAACL-HLT 2009
NAACL HLT 2009 Workshop on
Integer Linear Programming for Natural Language Processing

June 4, 2009, Boulder, Colorado, USA

Call for Papers
(Submission deadline: March 15, 2009)
Integer Linear Programming (ILP) has recently attracted much attention
within the NLP community.  Formulating problems using ILP has several
advantages.  It allows us to focus on the modelling of problems,
rather than engineering new search algorithms; provides the
opportunity to incorporate generic global constraints; and guarantees
exact inference.  This and the availability of off-the-shelf solvers
has lead to a large variety of natural language processing tasks being
formulated in the ILP framework, including semantic role labelling,
syntactic parsing, summarisation and joint information extraction.

The use of ILP brings many benefits and opportunities but there are
still challenges for the community; these include: formulations of new
applications, dealing with large-scale problems and understanding the
interaction between learning and inference at training and decision
time. The purpose of this workshop is to bring together researchers
interested in exploiting ILP for NLP applications and tackling the
issues involved. We are interested in a broad range of topics
including, but not limited to:

- Novel ILP formulations of NLP tasks.  This includes: the
introduction of ILP formulations of tasks yet to be tackled within the
framework; and novel formulations, such as equivalent LP relaxations,
that are more efficient to process than previous formulations.

- Learning and Inference.  This includes issues relating to:
decoupling of learning (e.g., learning through local classifiers) and
inference, learning with exact  (e.g., ILP) or approximate inference,
learning of constraints, learning weights for soft constraints, and
the impact of ignoring various constraints during learning.

- The utility of global hard and soft constraints in NLP.  Sometimes
constraints do not increase accuracy (and can even decrease it), when
and why do global constraints become useful?  For example, do global
constraints become more important if we have less data?

- Formulating and solving large NLP problems.  Applying ILP to hard
problems (such as parsing, machine translation and solving several NLP
tasks at once) often results in very large formulations which can be
impossible to solve directly by the ILP engine.  This may require
exploring different ILP solving methods (such as, approximate ILP
solvers/methods) or cutting plane and pricing techniques.

- Alternative declarative approaches.  A variety of other modeling
frameworks exist, of which ILP is just one instance.  Using other
approaches, such as weighted MAX-SAT, Constraint Satisfaction Problems
(CSP) or Markov Networks, could be more suitable than ILP in some
cases. It can also be helpful to model a problem in one framework
(e.g., Markov Networks) and solve them with another (e.g., ILP) by
using general mappings between representations.

- First Order Modelling Languages.  ILP, and other essentially
propositional languages, require the creation of wrapper code to
generate an ILP formulation for each problem instance. First (Higher)
Order languages, such as Learning Based Java and Markov Logic, reduce
this overhead and can also aid the solver to be more efficient.
Moreover, with such languages the automatic exploration of the model
space is easier.


We encourage submissions addressing the above questions and topics or
other relevant issues. Authors are invited to submit a full paper of
up to 8 pages (with up to 1 additional page for references), or an
abstract of up to 2 pages. Appropriate topics for abstracts include
preliminary results, application notes, descriptions of work in
progress, etc. Previously published papers cannot be accepted.

The submissions will be reviewed by the program committee. Note that
reviewing will be blind and hence no author information should be
included in the papers. Self-references that reveal the author's
identity, e.g., "We previously showed (Smith, 1991) …", should be
avoided. Instead, use citations such as "Smith previously showed
(Smith, 1991) …".

Papers will be accepted on or before 6 March 2009 in PDF format via
the START system at https://www.softconf.com/naacl-hlt09/ILPNLP2009/.
Submissions should follow the NAACL HLT 2009 formatting requirements
for full papers , found at

March 15, 2009: Submission deadline
March 30, 2009: Notification of acceptance
April 12, 2009: Camera-ready copies due
June 4, 2009: Workshop held in conjunction with NAACL HLT

INVITED SPEAKER: Dan Roth (University of Illinois at Urbana-Champaign)

- Dan Roth (University of Illinois at Urbana-Champaign)
- Mirella Lapata (University of Edinburgh)
- Scott Yih (Microsoft Research)
- Nick Rizzolo (University of Illinois at Urbana-Champaign)
- Ming-Wei Chang (University of Illinois at Urbana-Champaign)
- Ivan Meza-Ruiz (University of Edinburgh)
- Ryan McDonald (Google Research)
- Jenny Rose Finkel (Stanford University)
- Pascal Denis (INRIA Paris-Rocquencourt)
- Manfred Klenner (University of Zurich)
- Hal Daume III (University of Utah)
- Daniel Marcu (University of Southern California)
- Kevin Knight (University of Southern California)
- Katja Filippova (EML Research)
- Mark Dras (Macquarie University)
- Hiroya Takamura (Tokyo Institute of Technology)

- James Clarke (University of Illinois at Urbana-Champaign)
- Sebastian Riedel (University of Tokyo)

Email: ilpnlp2009 at gmail dot com
Website: http://www-tsujii.is.s.u-tokyo.ac.jp/ilpnlp/