Who are they talking about? Detecting mentions of social groups in political texts with supervised learning

Social groups' struggle for political power and representation is at the heart of political science theory. But quantifying which groups politicians' refer to, claim to represent, or address in their public communication presents researchers with challenges. Manual coding is costly and thus limits scalability. And while dictionaries enable automated measurement, it is difficult to compile lists of keywords that cover the whole universe of groups that politicians might mention and the various terms and phrases used to reference them. We present a novel supervised learning method for identifying group mentions in political texts that remedies these shortcomings. We first collect human annotations of a sample of sentences to determine the text passages that refer to social groups. We then use these annotated sentences to fine-tune a neural language model for contextualized word-level supervised classification. Applied to unlabeled texts, our method enables researchers to automatically detect and extract the variable-length text segments that contain group mentions in a reliable, resource-saving way. We illustrate the strengths of our method in a study of British and German parties' campaign communication and parliamentary speeches. These applications demonstrate that our method enables reliable retrieval of group mentions at scale, facilitates the inductive discovery of group mentions, and thus allows new quantitative insights into group-based political rhetoric.