Sentiment Stream: A Visualization Design Study

Text data, such as customer satisfaction reports, personal micro-blogs, or emails, is being produced like never before; but tools for visualizing this growing data source are lacking. Typical text visualizations, like word clouds and word trees, tend to leave out relevant context, remove dimensionality of the underlying data, and restrict access to source text. In this paper, we present the Sentiment Stream, a proposed form of text visualization. Sentiment Stream aggregates text documents by time and space bins, preserving temporal and spatial context, and finds the most emotional content in each bin using part of speech tagging and sentiment analysis. The most relevant word from each bin is displayed in a grid encoding time and space, sentiment is encoded using color, and objectivity is encoded using font weight. We demonstrate the application of Sentiment Stream on a dataset of 20 million Tweets following Hurricane Sandy in the week after landfall in October 2012 in New Jersey, USA. Because of the unique spatial, temporal, and emotional context provided by Sentiment Stream, we are able to make conclusions about the sentiment surrounding this natural disaster that would not have been possible using other text visualizations.