Fracking Twitter: Utilizing machine learning and natural language processing tools for identifying coalition and causal narratives

The Narrative Policy Framework (NPF) has provided policy scholars with a valuable method to gain empirical insight into the power of narratives in the policy process. However, a significant limitation of the NPF has been its ability to deploy this framework on large N datasets due to the labor-intensive nature of collecting narrative data. In recent years, NPF scholars have turned to computational social science tools to address this challenge. This study builds upon this emerging body of literature and our previous work, which uses sentiment analysis, a natural language processing technique, to evaluate the use of the angel/devil shift across coalitions before and after a major policy change. We examined Tweets that included the terms "fracking" and "New York" before and after the introduction of a moratorium. While sentiment analysis allowed us to gain insight into the narrative structure of the fracking policy discourse space, the labor involved in hand-coding Twitter users into neutral-, pro-, or anti-fracking groups was onerous. This project aims to supplement our natural language processing method by employing supervised machine learning techniques to increase the universe of respondents. We hand-coded 500 Twitter users into neutral-, pro-, or anti-fracking groups and trained a much larger dataset using an extreme gradient boost algorithm to classify a broader corpus of Tweets. This enabled us to expand the number of Tweets used in the analyses. We then applied sentiment analysis on this newly classified larger dataset to reveal differences in the pro-fracking and anti-fracking advocacy coalitions. By using machine learning to classify pro and con Tweets, we gained the ability to achieve significantly greater insight into how these two subgroups employed different narrative and linguistic devices in their Twitter discussions about fracking.