NIPS 2005 WORKSHOP : Advances in Structured Learning for Text and Speech Processing

This workshop is intended for researchers and students interested in developing and applying structured classification methods to text and speech processing problems. Recent advances in structured classification provide promising alternatives to the probabilistic generative models that have been the mainstay of speech recognition and statistical language processing. However, powerful features of probabilistic generative models, such as hidden variables and compositional combination of several kinds of evidence, do not transfer cleanly to all structured classification methods. Starting with surveys of the state-of-the-art in structured classification for text and speech, the workshop will focus on successes, failures, and directions for improvement of structured classification methods for text and speech and possible syntheses between the new structured classification methods and traditional generative models. Comparison will also be made with "generative" vs. "discriminative" training procedures in structure classification problems. A successful workshop will identify critical questions that current methods are not yet capable of solving, and promising directions for solution. For instance, we hope to achieve a better understanding of how discriminative models may work with missing information, such as under-specified alignments or syntactic analyses --- we plan, more generally, answer questions such as why, when, and where use a generative model. Such problems arise in both speech, language, and text processing, and will serve as unifying themes for the workshop.

Among questions to be discussed, we expect:

Discriminative vs. generative models and algorithms
Max margin, perceptron, and other criterion
Incorporating prior knowledge
Using data from multiple domains
Adaptation of structured classifiers to new conditions
Using unlabeled data
Combining text and speech
Integrated inference for complex language processing tasks

Organizers

Fernando Pereira	CIS, University of Pennsylvania
Michael Collins	CSAIL, MIT
Jeff Bilmes	EE, University of Washington
Koby Crammer	CIS, University of Pennsylvania

Program

This one-day workshop will have survey/tutorial talks in the morning followed by shorter contributed talks, posters, and discussion sessions later in the day. The survey talks will present central themes and questions that will guide the discussion during the workshop (see below). We are well aware of the tendency to turn workshops into mini-conferences, so we will make sure by keeping a tight control on the schedule that there will be sufficient time for discussion during and after talks. One of the ways we intend to use in order to accomplish this goal is to assign each of the presenters in the workshop to serve as discussant for someone else's presentation. Discussion will be moderated by the organizers.

The survey/tutorial talks are intended to provide a thorough background and overview of the field from a number of different perspectives (machine learning, statistics, mathematics, and applications such as speech, text, and language). In order to better customize the workshop to the interested audience, the survey/tutorial talks will be tuned to a set of issues and questions that are raised on a NIPS workshop discussion web page. The goal is for interested participants to post any nagging questions or general ideas that they have to this discussion board. These questions will then become a basis for the central theme of the workshop. Of course, for this to be a success it is necessary for people to pose questions to the discussion board. Therefore, it will be possible for people to post questions either with their identity associated, or anonymously. See below for further details.

Potential participants are encouraged to submit (extended) abstracts of two to four (2-4) pages in length outlining their research as it relates to the above theme. Papers may show novel ideas or applications related to structured classification. Encouraged topics include novel theoretical results, practical application results, novel insight regarding the above, and/or tips and tricks that work well empirically on a broad range of data. Papers should be formatted using the standard NIPS formatting guidelines.

Schedule and Dates

Nov 1st, 2005:
Paper submission deadline. Email all submissions ( .pdf file ) to: structlearn@lists.seas.upenn.edu with subject starting with 'STRUCTLEARN'.
Nov 8th, 2005:
Acceptance (talks and poster) decisions announced.
Dec 9th, 2005:
NIPS Workshop date.

Links

NIPS workshop web page
Discussion Board for Advances in Structured Learning for Text and Speech Processing
Please visit this web page and use it to post questions, problems, or ideas about open problems in the structured prediction area that you would like to see discussed both during the survey/tutorial talks and throughout the rest of the workshop.