Essays on Representation Learning for Political Science Research
Wu, Patrick
2021
Abstract
This dissertation consists of three papers about leveraging representation learning for political science research. Representation learning refers to techniques that learn a mapping between input data and a feature vector or tensor with respect to a task, such as classification or regression. These vectors or tensors capture abstract and relevant concepts in the data, making it easier to extract information. In the three papers, I show how representation learning allows political scientists to work with complex data such as text and images effectively. In the first paper, I propose using word embeddings to calculate partisan associations from Twitter users' bios. It only requires that some users in the corpus of tweets use partisan words in their bios. Intuitively, the word embeddings learn associations between non-partisan and partisan words from bios and extend those associations to all users. I apply the method to a collection of users who tweeted about election incidents during the 2016 United States general election. Which partisan accounts get retweeted, favorited, and followed, and which partisan hashtags are used closely correlate with the partisan association scores. I also apply the method to users who tweeted about masks during the COVID-19 pandemic. I find that users with more Democratic-leaning partisan association scores are more likely to use health advocacy hashtags, such as #MaskUp. In the second paper, I look at the automated classification of observations with both images and text. Most state-of-the-art vision-and-language models are unusable for most political science research, as they require all observations to have both image and text and require computationally expensive pretraining. This paper proposes a novel vision-and-language framework called multimodal representations using modality translation, or MARMOT. MARMOT presents two methodological contributions: it constructs representations for observations missing image or text, and it replaces computationally expensive pretraining with modality translation. Modality translation learns the patterns between images and their captions. MARMOT outperforms an ensemble text-only classifier in 19 of 20 categories in multilabel classifications of tweets reporting election incidents during the 2016 U.S. general election. MARMOT also shows significant improvements over the results of benchmark multimodal models on the Hateful Memes dataset, improving the best accuracy and area under the receiver operating characteristic curve (AUC) set by VisualBERT from 0.6473 to 0.6760 and 0.7141 to 0.7530, respectively. In the third paper, I turn to the issue of computationally studying language usage evolution over time. The corpora that political scientists typically work with are much smaller than the extensive corpora used in natural language processing research. Training a word embedding space over each period, the usual approach to studying language usage evolution, worsens the problem by splitting up the corpus into even smaller corpora. This paper proposes a framework that uses pretrained and non-pretrained embeddings to learn time-specific word embeddings, called the pretrained-augmented embeddings (PAE) framework. In the first application, I apply the PAE framework to a corpus of New York Times text data spanning several decades. The PAE framework matches human judgments of how specific words evolve in their usage much more closely than existing methods. In the second application, I apply the PAE framework to a corpus of tweets written during the COVID-19 pandemic about masking. The PAE framework automatically detects shifts in discussions about specific events during the COVID-19 pandemic vis-a-vis the keyword of interest.Deep Blue DOI
Subjects
computational social science representation learning natural language processing computer vision multimodal machine learning social media
Types
Thesis
Metadata
Show full item recordCollections
Remediation of Harmful Language
The University of Michigan Library aims to describe library materials in a way that respects the people and communities who create, use, and are represented in our collections. Report harmful or offensive language in catalog records, finding aids, or elsewhere in our collections anonymously through our metadata feedback form. More information at Remediation of Harmful Language.
Accessibility
If you are unable to use this file in its current format, please select the Contact Us link and we can modify it to make it more accessible to you.