Separating Fact from Fear: Tracking Flu Infections on Twitter

Proceedings of NAACL-HLT 2013, pages 789–795, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Separating Fact from Fear: Tracking Flu Infections on Twitter Alex Lamb, Michael J. Paul, Mark Dredze Human Language Technology Center of Excellence Department of Computer Science Johns Hopkins University Baltimore, MD 21218 {alamb3,mpaul19,mdredze}@jhu.edu Abstract Twitter has been shown to be a fast and reli- able method for disease surveillance of com- mon illnesses like influenza. However, previ- ous work has relied on simple content anal- ysis, which conflates flu tweets that report infection with those that express concerned awareness of the flu. By discriminating these categories, as well as tweets about the authors versus about others, we demonstrate signifi- cant improvements on influenza surveillance using Twitter. 1 Introduction Twitter is a fantastic data resource for many tasks: measuring political (O’Connor et al., 2010; Tumas- jan et al., 2010), and general sentiment (Bollen et al., 2011), studying linguistic variation (Eisenstein et al., 2010) and detecting earthquakes (Sakaki et al., 2010). Similarly, Twitter has proven useful for public health applications (Dredze, 2012), primar- ily disease surveillance (Collier, 2012; Signorini et al., 2011), whereby public health officials track in- fection rates of common diseases. Standard govern- ment data sources take weeks while Twitter provides an immediate population measure. Strategies for Twitter influenza surveillance in- clude supervised classification (Culotta, 2010b; Cu- lotta, 2010a; Eiji Aramaki and Morita, 2011), un- supervised models for disease discovery (Paul and Dredze, 2011), keyword counting1, tracking geo- graphic illness propagation (Sadilek et al., 2012b), and combining tweet contents with the social net- work (Sadilek et al., 2012a) and location informa- 1The DHHS competition relied solely on keyword counting. http://www.nowtrendingchallenge.com/ tion (Asta and Shalizi, 2012). All of these methods rely on a relatively simple NLP approach to analyz- ing the tweet content, i.e. n-gram models for classi- fying related or not related to the flu. Yet examining flu tweets yields a more complex picture: • going over to a friends house to check on her son. he has the flu and i am worried about him • Starting to get worried about swine flu... Both are related to the flu and express worry, but tell a different story. The first reports an infec- tion of another person, while the second expresses the author’s concerned awareness. While infection tweets indicate a rise in infection rate, awareness tweets may not. Automatically making these dis- tinctions may improve influenza surveillance, yet re- quires more than keywords. We present an approach for differentiating be- tween flu infection and concerned awareness tweets, as well as self vs other, by relying on a deeper analy- sis of the tweet. We present our features and demon- strate improvements in influenza surveillance. 1.1 Related Work Much of the early work on web-based influenza surveillance relied on query logs and click-through data from search engines (Eysenbach, 2006), most famously Google’s Flu Trends service (Ginsberg et al., 2008; Cook et al., 2011). Other sources of in- formation include articles from the news media and …

Content truncated to 500 words