As humans, we think of the social world in causal terms, seeking to answer questions that help us make sense of what happens around us, predict future outcomes, and address current issues. The way we communicate displays our causal reasoning, as it is through language that we share how we think the world works, and also attempt to control how others perceive who is responsible of social issues and how to fix them. Therefore, analyzing the use of causal expressions in the public sphere is crucial to better understand political discourse and its downstream effects on political opinion, attitudes, and behavior. In this dissertation, I develop a framework to identify and extract expressions of causal attributions from political text, drawing on techniques from computational linguistics. To do this, I design a coding scheme and create a corpus annotated for causal language in the political domain. I test the corpus on a broad range of pre-trained language models on the task of identifying and parsing expressions of causality: from smaller transformer-based models such as Bert, to modern Large Language Models, like GPT-4. Additionally, I introduce an evaluation strategy tailored to test if a model’s causal extraction is done through linguistic parsing or by leveraging causal patterns learned during pre-training. After interpreting experimental results, I develop a framework to parse expressions of causal attributions in text, adapted for political methodology. With this novel measurement tool, I empirically examine the constructions of causality used in the coverage of two international conflicts. I demonstrate the method’s capabilities and ease of use for political science research. With insights from the case study, I conceptualize “non attribution”, and posit it as a unique framing practice, complementary to responsibility framing. With it, I argue that the omission of causal attributions is a distinct frame, and not a neutral absence in news coverage. With this dissertation, I extend the current scholarship on causal language modeling, and show how new text-as-data methods can be used to scale and examine discourse to identify novel concepts, and extend previous work in political communication.
Paulina García-Corral (Thu,) studied this question.