Classifying Salient Textual Entities in the Headlines and Captions of Grouped Bar Charts

Information graphics, such as grouped bar charts, generally have a communicative message that they are intended to convey when they appear in popular media. Communicative signals are typically designed into the graphic to help convey to the graph viewer these intended messages. We have designed and implemented a system that automatically hypothesizes the intended message of a grouped bar chart from communicative signals that are automatically extracted from the graph. Analysis of our system revealed that textual evidence, such as graph entities mentioned in the headline or caption of the graphic, was the most important piece of evidence in our system. This paper describes a support vector machine classifier that takes a graph and its headlines and captions and predicts whether an entity is linguistically salient.