Text as Data

Summary

Text-as-data methods enable political scientists and public policy scholars to systematically analyze large collections of political documents—legislative speeches, agency reports, press releases, party manifestos, and court opinions—by treating text as quantitative data from which social scientific measurements can be extracted. These methods have transformed the study of legislative behavior, ideological positioning, bureaucratic communication, and policy framing by providing scalable, replicable alternatives to traditional hand-coding.

Three paradigms, distinguished by how researchers represent text computationally, organize this methodological landscape. The bag-of-words (BoW) paradigm treats each document as a collection of word frequencies, enabling transparent and interpretable tools—dictionary methods, scaling models such as Wordscores and Wordfish, and topic models such as latent Dirichlet allocation and structural topic model—that have become staples of political text analysis. The second paradigm—embedding—moves beyond word counts to represent text as dense vectors in semantic space, where meaning is captured through geometric relationships. Encoder-based transformer models (e.g., bidirectional encoder representations from transformers [BERT], sentence-BERT) power this paradigm, enabling semantic similarity measurement, cross-lingual comparison, and transfer learning that dramatically reduces annotation costs. The third, generative AI (GenAI), paradigm marks a qualitative shift: Decoder-based large language models read and interpret text through natural language prompts, performing analytical operations—from ideological scaling to document annotation—that no numerical representation can fully capture.

Rather than forming a linear progression in which newer methods supersede older ones, the three paradigms are best understood as complements. Using a common data set of U.S. Senate floor speeches, worked examples demonstrate how BoW methods reveal partisan word usage patterns, how embeddings capture semantic relationships invisible to word counts, and how zero-shot classification bridges toward the GenAI paradigm—showing convergent findings that strengthen substantive conclusions. Throughout, four established principles of text-as-data research serve as guideposts: All models of language are wrong, but some are useful; automation augments rather than replaces human judgment; no method is globally best; and validation is nonnegotiable.