Using communicative patterns to predict Twitter users' social capital, likability, and popularity gains with natural language processing

Abstract

Social media constructs a computer-mediated public space where individuals' visibility and influence can be quantitatively measured by the number of likes, retweets, and followers they receive. These metrics serve as a reward system that not only reflects users' popularity and social capital but also influences the climate of public opinion and deliberative democracy by encouraging and discouraging certain types of communication. Through analyzing Twitter data collected from U.S. congressional politicians and ordinary U.S. Twitter users in seven/eight waves, this study explores how communicative patterns--dual-process styles and sentiment--predict users' social capital, likability, and popularity gains on Twitter as well as how political identity and intergroup communication moderate the relationships between these variables. It found that: (a) rational expressions increase social capital and popularity gains while emotional expressions increase likability gains; (b) positive expressions generate a curvilinear effect on social capital, likability, and popularity gains in the politician dataset; (c) compared with Democratic users, Republican users receive relatively more social capital, likability, and popularity gains from emotional and negative expressions than from rational and positive expressions; (d) rational expressions lead to relatively more likability and popularity gains than emotional expressions in a group-salient context; and (e) positive expressions in ingroup/outgroup conversations generate opposite effects in the politician and ordinary user datasets. In addition, this study develops and advances computational methods in detecting communicative patterns, political identities, and intergroup communication. By implementing Distributed Dictionary Representations, this study creates metrics to measure dual-process thinking styles and sentiment in text; by developing a two-step model with deep learning using an attention mechanism, this study creates an interpretable method to detect political partisanship and intergroup communication.