2013 International Conference on Social Media and Society

Introduction: This paper describes the relationship between the language people use in tweets in a Twitter smoking cessation group and being successfully quit (abstinent) for 60 days. This paper uses a semantic network methodology by comparing the semantic network of tweets of abstinent individuals to the semantic network of non-abstinent individuals. The method presented here may be used for sentiment or content analysis in other online conversations, including contexts with large textual datasets. 

Sample: Adults in the United States who smoked at least 100 cigarettes in their lifetime and who smoked 5 or more cigarettes per day were included. Participants averaged 20 cigarettes/day (SD=9) and 18 years of smoking (SD=10). Participants were recruited online through Google AdWords, were asked to participate in one of two 12-week online Twitter support groups and were provided nicotine replacement therapy as an incentive for their participation (Ngroup1=20 with Ntweetsgroup1=1,125, Ngroup2=20 with Ntweetsgroup2=1,782). 

Methods: Two basic analyses were completed comparing the semantic network of tweets produced by 60-day abstinent participants and non-abstinent participants: 1) analyses comparing the actual terms used (frequency and betweenness centrality) and 2) descriptive analyses comparing the underlying network structure. Stopwords (e.g. the, at) were removed and a synonym set was created to match roughly equivalent words (e.g. lol and lolol). Further, a unique category was created for a number of concepts including: group member names, number of days of successful quitting and number of puffs. Thus, when a person posted the number of days smoke-free (days_free), this was treated as equivalent in the semantic network analysis to any other post about days smoke-free. Automap (Carley & Diesner, 2005) was used to create the semantic network data, which were then converted into adjacency matrices and analyzed using UCINet (Borgatti, Everett, & Freeman, 2002). 

Results: All networks were sparse (density=0.04-0.06). All networks contained discrete clusters with one primary cluster containing most nodes. Emotional terms were typically supportive (lol, great, good) and were frequent across all networks. Participant name (PN) was most frequent across three of the four semantic networks (except in group 2 among those abstinent, where PN was ranked 6). PN also tended to have the high-normalized Freeman's betweenness centrality (NBet) (which is a measure of the shortest path between terms) (group1abstinent=18.01, group2abstinent=3.86, group1nonabstinent=16.16, group2nonabstinent=6.99). The messages sent by participants who reported a 60-day successful quit attempt contained proportionately more terms about the 1) number of days smoke-free (NBetgroup1=8.63 and NBetgroup2=3.43) and quitting related words, e.g. quit (NBetgroup1=7.52 and NBetgroup2=5.96). The messages sent by participants who did not have a successful 60-day quit attempt were more likely to include smoking words like smoking (NBetgroup1=7.19 and NBetgroup2=7.43) and smoke (NBetgroup1=5.82 and NBetgroup2=5.94). 

Discussion: Results suggested a relationship between how people talk about smoking cessation in an online forum and successful 60-day abstinence. Future analyses should test the relationship between early use of these terms and 60-day abstinence. Further, results suggest that this type of semantic network analysis may be a viable method for analyzing content or sentiment in Twitter data. 

Borgatti, S. P., Everett, M. G., & Freeman, L. C. (2002). Ucinet for Windows: Software for Social Network Analysis. Harvard, MA:: Analytic Technologies. 
Carley, K. M., & Diesner, J. (2005). AutoMap: Software for Network Text Analysis.