Selecting More Informative Training Sets with Fewer Observations

Abstract

A standard text-as-data workflow in the social sciences involves identifying a set of documents to be labeled, selecting a random sample of them to label using research assistants, training a supervised learner to label the remaining documents, and validating that model's performance using standard accuracy metrics. The most resource-intensive component of this is the hand-labeling: carefully reading documents, training research assistants, and paying human coders to label documents in duplicate or more. We show that hand-coding an algorithmically selected rather than a simple-random sample can improve model performance above baseline by as much as 50%, or reduce hand-coding costs by up to two-thirds, in applications predicting (1) U.S. executive-order significance and (2) financial sentiment on social media. We accompany this manuscript with open-source software to implement these tools, which we hope can make supervised learning cheaper and more accessible to researchers.