Dissertation: Essays on Computational Technologies in International Relations
Advisors: Austin Carson (Chair), James Evans, Rochelle Terman
My dissertation explores how emerging computational technologies—encompassing machine learning tools, the data that powers them in commercial settings, and the consumer services and devices these tools are designed to optimize—are theoretically and methodologically relevant for the field of international relations.
Theoretically, I argue that the world’s largest technology companies are critical actors in international politics due to their ownership of powerful data collection and processing infrastructures. By developing and deploying various computational technologies, technology firms set the terms by which an immensely differentiated world is commonly understood. This allows firms to shape the public perception of political issues and perform functions akin to those of governments.
Methodologically, I argue that researchers can use these same computational technologies to offer novel insights across diverse debates in political science. I demonstrate this by employing natural language processing and computer vision methods to address both historical and contemporary debates in the field.
Taking the form of a three-paper dissertation, the first paper presents my main theoretical contribution. The second paper leverages embeddings of geolocated Google search data to measure how Google differentially portrays political information across locations and languages. The third paper uses word embedding models to measure identity and friendship among countries in manners which align with constructivist theories of international relations.
Auditing Google Search Results for Human Rights
With Dr. Rochelle Terman (Under Review)
Research on international norms emphasizes the role of information, but often relies on a key premise: that the messages produced by international actors reach intended audiences. Yet this assumption is rarely studied directly, leaving us with scant knowledge about what people actually learn when accessing information about norms like human rights. This research note reports findings from an audit of Google search for human rights queries across 146 countries and 70 languages. We find that search results vary markedly across countries, indicating that different users are directed to different resources about human rights. Some people receive information produced by prominent international organizations, while others learn about human rights mainly from governments and local publishers. These disparities likely reflect differences not only in the sources of content, but also in the substantive messages about human rights. Furthermore, we find that natural language—far more than any other factor—correlates strongly with variation in search results, suggesting a prominent “language barrier” to international human rights discourse. Our results have significant implications for debates surrounding norm diffusion and knowledge production in world politics, while offering practical guidance for human rights advocates.
Measuring Perceptions of the Societal Impact of Social Media with Text Generation Models
Researchers regularly use surveys to gather public opinion data, yet a primary limitation of survey research lies in its confinement to the current time frame. To evaluate shifts in public opinion over time, researchers must repeatedly conduct survey waves, a task that can be both time-consuming and costly. This paper introduces a novel survey-style research approach that involves fine-tuning large language models on subsets of temporally and ideologically distinct social media data. In essence, I fine-tune a group of GPT models with Reddit data, each representing a specific political ideology at a particular time, and subsequently prompt these models with a set of survey questions. I pilot this methodology to examine public perceptions of the societal role of social media in the United States from 2017 to 2019. The findings provide suggestive evidence that liberals more so than conservatives grew to view social media, particularly Facebook, as having a greater impact on society during this period. Notably, these views experienced the most significant changes following the public revelations of Facebook's Cambridge Analytica scandal in 2018.
Text As Data
With Dr. Rochelle Terman, in The Handbook of Research Methods for International Relations
This chapter evaluates a variety of strategies for computational text analysis, otherwise known as "text as data" (TAD). We begin by providing general guidance on constructing a corpus – or a collection of documents – relevant to a research question, followed by an explanation of the necessary steps to prepare raw text for a variety of computational models. We then describe three TAD modeling strategies. First, we discuss dictionary methods, which leverage the frequency and tone of words to measure documents on a quantity of interest and/or classify them into a particular category. Second, we review supervised classification, where a machine learning algorithm assigns documents to predetermined categories. Finally, we address unsupervised techniques, which seek to discover new ways of organizing texts that are theoretically useful, but perhaps understudied or previously unknown.
Navigating the Rift: Firm Strategic Response to International Events
What is the connection between international political tensions and economic relations? Do political events between governments prompt changes in firm behavior? I argue that despite the necessary sunk costs of conducting business in a globalized world, a firm’s international activity is shaped by political events between states. Using firm-level global revenue exposure data from FactSet Research Systems and the Global Database of Events, Location, and Tone (GDELT), I find evidence that: 1) firms seek to conduct more business in adversarial states, possibly as a means to hedge against the prospect of future regulation, 2) firm size is an important determinant of firm response to international political events, and 3) events involving military actors are of greater consequence to firms than events involving government actors.
This paper served as my masters thesis at UChicago.