Models of natural language processing problems are often incredibly complex, and there is never enough data to properly estimate all the required parameters. This has lead to a strong need for learning techniques with built-in capacity control; most classical solutions to this problem involve largely ad-hoc smoothing techniques. The application of Bayesian learning methods to these problems could potentially result in more effective models, for which extensive cross-validation is no longer required for hyperparameter tuning or model selection.

The goals of this workshop are to bring together researchers from both the Bayesian machine learning community and the natural language processing community to enable cross-fertilization of techniques, models and applications. We wish to focus on the following issues:

• Statistical Models: Current Bayesian models for text have largely focused on "bag of words" style approaches, where conditional independence is assumed between words. This leads to a convenient interpretation of a document as a sequence of draws from multinomial distributions, but does not account for any of the internal structure that exists in documents and which NLP researchers are interested in. How can we build models that move beyond the bag of words assumption? What structures are useful for modeling? How can we model these structures efficiently? Can we learn these models automatically?
  
  
• Applications-oriented Models: Many statistical models for text have aimed at automatically inferring implicit relationship between varied elements of documents in a corpus. How can we use such models to aid in applications? Can we develop similar models that are aimed at solving a real-world NLP task? For what NLP applications are Bayesian techniques appropriate and how can we develop models specific to these problems?

• Kenneth Church, Microsoft Research
• Tom Griffiths, Brown University
• Andrew McCallum, University of Massachusetts Amherst

• Kenneth Church, Microsoft Research
• Tom Griffiths, Brown University
• Jeff Bilmes, University of Washington
• Andrew McCallum, University of Massachusetts at Amherst
• Sam Roweis, University of Toronto

Please see the call for participation for more information on how to submit workshop papers. Also, if you are a researcher in either the Bayesian learning community or the NLP community, please complete our survey, which will serve to guide the panel discussions at the workshop.