Panelist: Andy Kehler
Title: Opportunities and Challenges for a Bayesian Approach to Language Processing 

Computational approaches to language understanding are typically
reactive: language input triggers a search for an interpretation.
Human language interpretation, on the other hand, is proactive:
comprehenders use context to create 'top-down' expectations about the
ensuing message and integrate them with the 'bottom-up' evidence
provided by the speaker's utterance.  Bayes' Rule naturally captures
these two contributors to interpretation via its prior and likelihood
terms respectively.  I will discuss recent research (joint with Hannah
Rohde) that reveals that human pronoun interpretation behavior not
only follows Bayesian principles, but that much of the complexity that
affects interpretation resides in contextual factors that condition
the prior: that is, the part of the equation that is independent of
the speaker's decision to use a pronoun.  These results open up new
possibilities for training systems, since large corpora of annotated
pronouns should not be necessary to capture these complexities.  At
the same time, they also reveal the importance of overcoming
significant uphill battles associated with predicting meaning that
context implicitly conveys.