Understanding how people communicate during disasters is important for creating systems to support this communication. Twitter is commonly used to broadcast information from the ground and to organize support during times of need. During the 2010 Gulf Oil Spill, Twitter was utilized for spreading information, sharing firsthand observations, and to voice concern about the situation. Through building a series of classifiers to detect emotion and sentiment, the distribution of emotion during the Gulf Oil Spill can be analyzed and its propagation compared against released information and corresponding events. I contribute a series of emotion classifiers trained from 4,000 tweets with preliminary results of accuracy between 51% and 70%. These classifiers are used to analyze the emotional impact of events, such as Obama’s Oval Office Address. A broader implication of this project includes analysis of other disaster datasets through utilizing our open-source visualization and classifiers.

This project – an ongoing collaboration between the Emerging Capacities of Mass Participation (emCOMP) and Social Media (SoMe) Laboratories at UW – investigated the spread of misinformation during the Boston Marathon Bombings over Twitter, ultimately leveraging social media in crisis situations for communication, news reporting, and aid coordination efforts. Thus far we classified six rumors within the roughly twenty million tweet dataset, isolated several temporal “signatures” representative of certain kinds of rumors, and identified tweet characteristics (URLs and location data, for instance) associated with either misinformation or correction. We then further investigated the role of URLs and domains in the spread of misinformation, postulating that outside sources play a significant role in the development of temporal signatures. Finally, using a combination of rumor classification and machine learning, we began to create an application that automatically identifies misinformation in real time as it spreads.