Authors: Arif Khan, Shahadat Uddin and Nazim Choudhury
Ebola outbreak, one of the major events of 2014, has claimed over 10,000 lives (CDC, 2015) so far. Although started earlier, it gained worldwide attention from September, 2014 when the first case was confirmed in United States (Patwardhan, 2014). Over the next few months’ window, the topic kept dominating in social networks (Househ, 2015) with varying responses. Apart from sharing knowledge and awareness, there was also widespread sentiment of panic, criticism and satire that propagated throughout the platform. Given the complex and longitudinal nature of data, it is therefore important to explore the dynamics of this particular event to understand how people’s sentiments evolve as they share and re-share information in times of crisis like Ebola.
Objective: This research focuses on understanding the evolution of sentiments following the events of Ebola crisis. We aim to explore this dynamics from three perspectives. Firstly, from actor’s perspective where we like to identify major contributors (e.g., news agency, health organizations and individuals) of information from social network perspectives as well as their structural position within the network. Secondly, we are interested in the longitudinal analysis of sentiment propagation i.e., how fast different sentiments diffuse through the network and the time lag between an actual event occurrence and the time when that reaches most of the followers. Finally, we want to explore spatial impact on sentiments e.g., how they vary within different states or countries.
Methods: We chose Twitter as social network platform. Related tweets were downloaded with custom software that utilizes web mining and Twitter APIs. We chose three months’ window (September – November, 2014) when related events and responses (i.e., confirmed cases, patient transfers, quarantine and deaths in Europe and United States) were at the peak outside of western Africa (Times, 2015). Data were segmented into entities e.g., text, hashtags, user, geolocation etc. and saved into database. As an ongoing work, we are using sentiment analysis software to identify subjective information. Next, we will use social network analysis methods like centrality and clustering algorithms to find prominent groups or actors within the network. We also applied a set of measures proposed by Uddin et al. (2015) to understand the actor level dynamics of longitudinal network. These measures can identify dynamicity of different sentiments over time.
Results: The database consists of approximately 1.56 million tweets from 0.6 million users over 3 months. Half of the tweets are from United States followed by U.K., Brazil and Spain. Significant amount of information originated from dedicated accounts on Ebola crisis followed by general media agencies, healthcare organizations and individuals. Although sentiment analysis is still undergoing, initial result suggests that significant proportion of popular tweets have witty or criticizing (related to politics or celebrities) tone rather than sharing fact or information.
Future Work: We are still applying sentiment analysis to quantify different polarities of tweets. After that, social network based structural and dynamicity measures will be applied. Furthermore, the result should be normalized to remove any bias because not every tweet is made available via Twitter API.
