Discovering evolving political vocabulary in social media

Abstract

As a surrogate data source for many real-world phenomena, social media such as Twitter can yield key insight into people's behavior and their group affiliations and memberships. As an event unfolds on Twitter, the language, hashtags, and vocabulary used to describe it evolves over time, so that it is difficult to a priori capture the composition of a social group of interest using static keywords. Capturing such dynamic compositions is crucial to both understanding the true membership of social groups and in providing high-quality data for downstream applications such as trend forecasting. We propose a novel unsupervised learning algorithm that builds dynamic vocabularies using probabilistic soft logic (PSL), a framework for probabilistic reasoning over relational domains. Using 10 presidential elections from eight countries of Latin America (Mexico, Venezuela, Ecuador, Paraguay, Chile, Panama, Colombia, and Honduras), we demonstrate how our vocabulary-discovery approach helps capture dynamic trends specific to each election. The ability to grow a vocabulary concurrently with social media trends helps capture key milestones in election campaigns.