October 24, 2017
Dhiraj Murthy and his colleagues are sifting through the cacophony of social media, during Hurricane Harvey, for signals we can use to save lives.
By Daniel Oppenheimer
Although 9-1-1 systems remain at the core of emergency response in the midst of disasters like hurricanes Harvey and Irma, it is becoming clear to emergency responders and policymakers that social media has emerged as an alternative emergency response infrastructure for when 9-1-1 is overloaded or unavailable. People are asking for help on Twitter and Facebook and Instagram, and both official emergency responders and volunteers are going to these platforms to figure out who to help and where.
One of the obvious questions that follows from this fact, for people who need to think about these things, is how to optimally sift through the vast amount of information on social media for useful information to guide emergency response.
This is the question that Dhiraj Murthy, an Associate Professor in the School of Journalism and the Department of Sociology at The University of Texas at Austin, is now working to answer, with a grant from the National Science Foundation. By analyzing the ways in which victims and rescuers use Twitter, Facebook, Instagram, YouTube, and other social media channels in crisis, Murthy and principal investigator Keri Stephens hope to be able to develop scalable methods to improve current and future resilience during weather emergencies.
Texas Health Journal spoke to Dr. Murthy about this work, and about the challenges of accessing and analyzing the vast but not always publicly available universe of social media data.
Murthy is director of the Computational Media Lab. His research explores social media, virtual organizations, virtual teams, digital research methods, race/ethnicity, and big data quantitative analysis.
Texas Health Journal: What are some of the important questions when it comes to social media and emergencies?
Dhiraj Murthy: We tend to associate emergency health lifelines with 9-1-1. In extreme weather emergencies, however, 9-1-1 can become overloaded. It can be challenging or even impossible to get access to critical, lifesaving care, or emergency evacuation, through 9-1-1. As we saw in the cases of Hurricanes Harvey and Irma, disaster victims use their smartphones. But how are they using them? How can first responders better understand this lifeline? Unlike 9-1-1, social media hosts non-emergency and emergency content simultaneously during disasters, so it is a potential resource for emergency response but only if can we discern the useful signal from the noise to ‘hear’ what victims are saying or have said during a weather emergency. Ultimately, can we minimize injury and loss of life during weather emergencies if we better understand social media use during these events?
Are there situations in which it’s better for an individual to ask for help on Twitter or Facebook, say, than to call 9-1-1? Does it depend on the nature of the emergency? During the flooding in Houston, for instance, the help wasn’t coming exclusively from first responders. It was also coming from fellow Houstonians, and other civilians, who just went out with their boats to help.
Absolutely. The situations that come most to mind are (1) when 9-1-1/emergency services are overloaded and/or (2) when it would take too long for official first responders to reach an individual.
The nature of the emergency is also important. If the weather emergency is of a particular scale and volunteer organizations and individuals are involved, social media may play a far greater role. In the case of Houston, the sheer level of damage was not only highly unexpected, but contingency resources were not always quickly available. Volunteers, however, snapped into action. We saw individuals relying on public Facebook pages or checking Twitter hashtags for information about who and where to help, and then volunteering with their boats to go out and rescue others. Social media platforms afforded possibilities for both centralized and decentralized volunteer rescue efforts. In other words, social media such as Facebook and Twitter provided communications infrastructure that enabled volunteer organizations and citizen rescuers to assist Harvey victims, even though these are resources not connected to the formal networks behind 9-1-1.
How do you perform your analyses of what’s happening on social media? I’ve always been struck by the fact that Google, for instance, doesn’t seem to be great at picking up what’s happening on Twitter or Facebook. It seems better with blogs, but even then not that great. So where’s the data for you to analyze? I can’t imagine the big social media companies are just making it accessible to researchers.
Understanding anything that's happening on social media presents challenges.
One challenge is data collection. Social media companies are private entities driven primarily by shareholder profit. Though researchers would benefit from unfettered access to their data, we have quite limited access. In the case of Twitter, for example, researchers have access to 1% of all real-time tweets via Twitter’s streaming Application Programming Interface (API), the public gateway to Twitter’s servers. Other data has to be paid for, making scaling challenging.
Second, even if one has access to these data, understanding social media data is like drinking from a firehose. These data have a lot of ‘noise,’ making it difficult to find the signals coming from victims during a disaster. For example, most social media-based research regarding weather emergencies collects data using catchall keywords or hashtags such as #Harvey or #Hurricane. In our research, we have decided to flip this method, and to first speak to victims in Houston and Florida to understand what types of content they were actually posting and what typical ‘calls’ for help on social media actually looked like.
The purpose of our work is to develop an ontology that is built from knowledge gained from interviews and surveys to be done with Harvey and Irma victims. The idea here is to identify social media channels and the specific attributes of posted content in order to be able to inform 9-1-1 and other official first responder stakeholders. We will then collect data retroactively, but via highly targeted methods. We expect to provide decision-makers with useful guidance in the very short term, within our one-year granting period, to help improve health outcomes in future emergencies.
That is fascinating. So, if I understand correctly, the goal is ultimately to give emergency responders tools and guidelines for how, in the midst of a crisis, to sift through social media more efficiently for useful information. Is that right? Are we talking about a set of hashtags they should flag? Or a sequence of words? Or certain kinds of images? Or even more sophisticated algorithms that will somehow select out key tweets and posts?
Yes, we are trying gain knowledge about how people actually called for help on social media platforms rather than be led by a purely data-driven approach, which is the method that tends to be deployed in these contexts.
This knowledge would be translated into formal ontologies. By this, I mean emergency responders and policymakers would be given a relevant ontology that provides them with mechanisms to be able to identify these calls for help. Hashtags or sequences of words or certain kinds of images will definitely be included. However, there could be platform differences between Instagram, Facebook, YouTube, etc. We also do not actually know if certain platforms were used, such as YouTube. The idea is that the fieldwork would help us identify the venues that victims were going to and volunteers were using and be able to not only discern attributes such as sets of hashtags, sequences of words, certain kinds of images, but also differences in how the platforms were used, whether rescuers saw the content, and how volunteer rescuers saw the content as well. In other words, part of this is to identify what calls for help were more successful (i.e. heard) on social media as well as which ones were not heard. Therefore, there is also value in building ontologies around calls for help generally as well as specific attribute classes for heard calls and unheard calls so that those who were unheard could be more actively listened to in future weather emergencies. In addition, these individuals may still be in need of assistance.