Sentiment Analysis of Occupy Wall Street Tweets

THE rise of social media has changed political discourse around the world by extending the potential reach of otherwise marginal voices and creating opportunities for massive group coordination and action. The ”Arab Spring” protest movements of the past year demonstrate the political power of these tools. A survey by a news organization in the region reported that the “vast majority of . .. people surveyed over three weeks in March said they were getting their information from social media sites (88 per cent in Egypt and 94 per cent in Tunisia)” [5]. As the Twitter stream becomes richer in information and more significant in political impact, the value of monitoring and understanding the stream increases. The United States’ Department of Homeland Security, for example, recently announced its intention to develop a system for collecting intelligence information from Twitter and Facebook [2]. The immense quantity of the data available makes identifying key trends non-trivial, however, as Department of Homeland Security Undersecretary Carolyn Wagner comments: “We’re still trying to figure out how you use things like Twitter as a source . . .How do you establish trends and how do you then capture that in an intelligence product?” Human observers, inundated by thousands of tweets, need computational tools to aid them in analysis of social media data. While application programming interfaces allow easy automated collection of large social media datasets, analysis of these datasets remains difficult. Simple keyword searches can help identify topics discussed in social media posts, but much of the value of these data lies in the posters’ opinions about these topics, encoded in their messages but difficult to extract computationally. Fortunately, machine learning offers a potential solution through the field of sentiment analysis. A sentiment analysis algorithm “seeks to identify the viewpoint(s) underlying a text span” [8] by extracting descriptive features from text fragments and using them as inputs to a learned hypothesis function. Such algorithms have already been used to classify opinions on current events as expressed in news sources [4]. Asur and Huberman applied sentiment analysis to Twitter data to forecast box-office revenue for movies with competitive accuracy [1]. By training such an algorithm to recognize specific political sentiments of interest rather than opinions about movies, an observer could “predict the future” of relevant political movements much as Asur and Huberman predicted market behavior [1]. In this investigation we apply machine learning methods to analysis of political sentiment in social media. We concentrate on a specific political phenomenon, the Occupy Wall Street movement, and a particular social media platform, Twitter.com. Furthermore, we restrict ourselves to classifying postings into three categories: pro-OccupyWall-Street, anti-Occupy-Wall-Street, and neutral or unrelated. We apply a series of machine learning techniques to these data.