Abstract: This talk will be interactive. I will first briefly introduce
my research on the recognition of high-level messages in grouped bar
charts using a Bayesian methodology. Message categories are defined
based on an analysis of a grouped bar chart corpus containing multimodal
documents from popular media. The Bayesian network hypothesizes the most
likely message category for a given graphic, given the communicative
signals in the graphic. One communicative signal this framework does
include is the shallow parsing of the caption of the graphic. However,
it does not consider the article text of a multimodal document. I will
present a few ideas about how to automatically extract signals from the
article text, which will then be available to the Bayesian network as
evidence for certain high-level messages. However, this work is not
refined, as work typically is in a SIG-NLP talk. Instead, I will present
my insights about the positives and negatives about my approaches, and
then ask for comments about my ideas and also be open to suggestions
about other approaches.