Incremental Parsing: Bringing Engineering and Cognition Together
Workshop at ACL-2004
Barcelona, Spain, July 25, 2004
Workshop Topic
Much recent parsing research has focused on the limited task of
achieving broad coverage and high accuracy in parsing Treebank
corpora. The parsing models developed for this task typically work on
a sentence-by-sentence basis: they often only deliver a valid analysis
if the input consists of a complete sentence. They are not designed to
operate incrementally, i.e., to deliver partial analyses (perhaps with
associated probabilities) that can be updated on a word-by-word basis
as more of the input becomes available.
Incrementality is desirable for two reasons. First, incremental
processing is crucial for many NLP tasks. Language modeling, for
instance, typically requires that probabilities are assigned
incrementally as more and more of the speech stream becomes
available. Recently, a number of parsing models have been proposed
that have this property and thus can be used for language
modeling. These models have resulted in lower perplexity scores and
word error rates than the standard n-gram models. However, the parsing
accuracy of these models typically falls short of the state of the
art. The challenge for parsing research is to develop models that
achieve optimal performance for both parsing and language modeling.
The second argument for incrementality comes from cognitive
modeling. There is substantial evidence showing that humans process
language in an incremental fashion. Any cognitively plausible model of
human parsing must take incrementality into account, and the modeling
literature contains considerable discussion on the relevant
computational mechanisms. Recently, a number of models of human
parsing have been proposed that are based on computational linguistic
approaches, such as PCFGs and related statistical models, suggesting a
potential synergy between cognitively and technologically motivated
parsing research.
Target Audience
The aim of the workshop is to address the dual challenge of defining
incremental parsing models that are useful for engineering tasks such
as language modeling, while also contributing to our understanding and
modeling of the human parsing mechanism. The workshop will bring
together parsing researchers from the computational linguistics and
cognitive modeling communities, and we expect extensive
cross-fertilization from this interaction. From the computational
linguistic perspective, cognitive modeling presents new challenges for
parsing research, including new evaluation measures that go beyond
traditional parseval measures. On the other hand, computational
linguistics can contribute crucial methodological advances to
cognitive modeling. For instance, the application of probabilistic
parsing algorithms to cognitive tasks has important implications for
the recent debate on the role of frequency information in human
parsing.
Areas of Interest
Possible topics for workshop submissions include:
- architectures, methods, and algorithms for incremental parsing;
including symbolic, probabilistic, connectionist, and hybrid models
- applications of incremental models to parsing, language modeling,
and cognitive modeling
- evaluation using standard metrics (parseval, perplexity, word error
rate)
- evaluation against behavioral data (reaction times, eye-tracking
data, linguistic judgments)
- applications of incremental parsing models in computational
linguistics
Workshop Organizers
Frank Keller, University of Edinburgh
Stephen Clark, University of Edinburgh
Matthew Crocker, Saarland University
Mark Steedman, University of Edinburgh
Keynote Speakers
Brian Roark, Oregon
Graduate Institute, Oregon Health and Science University
Patrick Sturt,
Department of Psychology, University of Glasgow
Program Committee
Steve Abney, University of Michigan
Thorsten Brants, Google
Eugene Charniak, Brown University
Ciprian Chelba, Microsoft Research
Michael Collins, MIT
Jeffrey Elman, UCSD
Ted Gibson, MIT
John Hale, Michigan State University
Mark Johnson, Brown University
Gerard Kempen, University of Leiden
Oliver Lemon, University of Edinburgh
Massimo Poesio, University of Essex
Stefan Riezler, Palo Alto Research Center
Brian Roark, Oregon Health and Science University
Douglas Roland, UCSD
Ed Stabler, UCLA
Suzanne Stevenson, University of Toronto
Patrick Sturt, University of Glasgow
Contact Information
The organizers can be contacted at:
School of Informatics
University of Edinburgh
2 Buccleuch Place
Edinburgh EH8 9LW, UK
phone: +44-131-650-4407
fax: +44-131-650-4587
email: acl04_workshop@inf.ed.ac.uk
created 2004-01-05, last modified 2004-06-05