Cooooooooooooooollllllllllllll!!!!!!!!!!!!!! Using Word Lengthening to Detect Sentiment in Microblogs

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 562–570, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Cooooooooooooooollllllllllllll!!!!!!!!!!!!!! Using Word Lengthening to Detect Sentiment in Microblogs Samuel Brody School of Communication and Information Rutgers University sdbrody@gmail.com Nicholas Diakopoulos School of Communication and Information Rutgers University diakop@rutgers.edu Abstract We present an automatic method which lever- ages word lengthening to adapt a sentiment lexicon specifically for Twitter and similar so- cial messaging networks. The contributions of the paper are as follows. First, we call at- tention to lengthening as a widespread phe- nomenon in microblogs and social messag- ing, and demonstrate the importance of han- dling it correctly. We then show that lengthen- ing is strongly associated with subjectivity and sentiment. Finally, we present an automatic method which leverages this association to de- tect domain-specific sentiment- and emotion- bearing words. We evaluate our method by comparison to human judgments, and analyze its strengths and weaknesses. Our results are of interest to anyone analyzing sentiment in microblogs and social networks, whether for research or commercial purposes. 1 Introduction Recently, there has been a surge of interest in sen- timent analysis of Twitter messages. Many re- searchers (e.g., Bollen et al. 2011; Kivran-Swaine and Naaman 2011) are interested in studying struc- ture and interactions in social networks, where senti- ment can play an important role. Others use Twitter as a barometer for public mood and opinion in di- verse areas such as entertainment, politics and eco- nomics. For example, Diakopoulos and Shamma (2010) use Twitter messages posted in conjunction with the live presidential debate between Barack Obama and John McCain to gauge public opinion, Bollen et al. (2010) measure public mood on Twitter and use it to predict upcoming stock market fluc- tuations, and O’Connor et al. (2010) connect pub- lic opinion data from polls to sentiment expressed in Twitter messages along a timeline. A prerequisite of all such research is an effec- tive method for measuring the sentiment of a post or tweet. Due to the extremely informal nature of the medium, and the length restriction1, the lan- guage and jargon which is used in Twitter varies sig- nificantly from that of commonly studied text cor- pora. In addition, Twitter is a quickly evolving domain, and new terms are constantly being intro- duced. These factors pose difficulties to methods designed for conventional domains, such as news. One solution is to use human annotation. For exam- ple, Kivran-Swaine and Naaman (2011) use manual coding of tweets in several emotion categories (e.g., joy, sadness) for their research. Diakopoulos and Shamma (2010) use crowd sourcing via Amazon’s Mechanical Turk. Manual encoding usually offers a deeper understanding and correspondingly higher accuracy than shallow automatic methods. However, it is expensive and labor intensive and cannot be ap- plied in real time. Crowd-sourcing carries additional caveats of its own, such as issues of annotator exper- tise and reliability (see Diakopoulos and Shamma 2010). The automatic approach to sentiment analysis is commonly …

Content truncated to 500 words