Word Embeddings as Historical Public Opinion Proxies, 1854–1955

• In Event: Comparative-Historical Sociology Section Roundtables
  In Refereed Roundtable (60 minute): Comparative-Historical Roundtable 7

Abstract

Public opinion toward the U.S. government—especially trust and confidence—has declined steadily since the 1970s, yet the longer-run historical dynamics of these attitudes remain difficult to explain because systematic survey data are sparse prior to the mid-20th century. To address this deficit, this study develops a text-based approach to estimating historical public attitudes toward the U.S. president from 1854 to 1955. Using the American Stories archive of local newspapers, I train yearly word2vec models and construct attitudinal scales—such as approval and perceived power—by projecting president-specific word vectors onto concept dimensions defined by opposing sets of anchor terms. The validity of these measures is assessed in three ways. First, I regress the embedding-derived measurements with real-world data measuring aspects of the presidency, such as presidential approval and executive order usage, showcasing strong correlations between the two. Second, averaging attitudinal scores across presidential terms showcases patterns that align with established historical narratives, including the expansion of executive authority around the turn of the twentieth century and the shifting reputations of individual presidents. Lastly, changepoint detection identifies distinct attitudinal regimes that coincide with major political and economic transformations, such as the Civil War, the Spanish American War, the Great Depression, and World War II. Together, the findings suggest that large-scale textual data can recover historically meaningful signals about public sentiment. While the approach cannot disentangle changes in public attitudes from shifts in journalistic style and is limited to the American case, it demonstrates the potential of word embeddings to extend the study of public opinion into eras where polling did not yet exist.

Author

• Sebastian Preising, University of California-Los Angeles