Each node is a group of sentences discussing rationality in a similar way, identified within a given decade. Node size reflects how many sentences belong to that group. Colors indicate broader thematic clusters that persist across decades — nodes sharing the same color address rationality in comparable ways across time. See the paper for the full methodology. Two visualization modes are available. Static mode positions clusters using a force-directed layout that reflects overall semantic proximity. Temporal mode groups clusters by decade along the x-axis (1900–1919, 1920–1929, …, 2000–2009), with thematic clusters separated along the y-axis, making it easier to follow how topics evolve over time.
Click a node in the network to explore its content.
TF-IDF highlights words that are specific to a group of data in comparison to the entire dataset. In this context, it emphhasiezs terms specific to the sentences within the selected cluster in comparison to all sentences across clusters.
Sentences assigned to this cluster, ranked by semantic proximity. Two metrics of similarity are provided: (1) similarity to the cluster centroid gives the cluster sentences most representative of the cluster as a whole; (2) similarity to the representative vector gives cluster sentences most representative of rationality discussion at this period. You can also filter to only show sentences mentioning 'rational' or 'rationality' (since not all sentences in the cluster may explicitly mention these terms).
Articles with the most sentences in this cluster.
Most cited references by articles which have sentences in this cluster. The user can filter to only show references with a known Web of Science ID.
Each node is an article. Articles are connected when they cite similar references — a proxy for shared intellectual background. Colors indicate communities of articles that recurrently cluster together across consecutive 8-year time windows. Uncolored nodes belong to communities that are too small or too short-lived to be considered substantive. The spatial layout positions articles with similar citation profiles in close proximity. See the paper for the full methodology. Use the time window selector to explore how the citation structure changes across periods.
Click a cluster label to explore its content.
Share of articles per cluster within the selected time window.
TF-IDF highlights terms that are specific to a cluster compared to all articles across all time windows.
Sentences most representative of the cluster content for the period.
Most cited references among articles in the selected cluster.
Origins show where articles came from (t-1), destinies show where they go (t+1).
This article offers a concrete, method-driven demonstration of how unsupervised quantitative methods can enrich the history of economic thought. Focusing on the long and shifting history of rationality in economics, we assemble and analyze the most extensive English-language corpus ever used for a historical study of economics, with nearly 290,000 full-text journal articles published between 1900 and 2009, paired with structured citation data. Combining large language model–based semantic analysis with bibliometric and network methods, we trace how economists have discussed, reformulated, and contested rationality over more than a century. Our approach identifies sentences and articles most closely associated with rationality, groups them into semantic clusters and bibliometric communities within short time windows, and then aggregates these groupings over time. This multi-scale design makes it possible to both “zoom out” to capture broad intellectual transformations and “zoom in” to examine specific debates, research programs, and moments of reception. Beyond substantive findings—illustrated through the contrasting trajectories of bounded rationality and behavioral economics—the article advances a broader methodological argument. We show that unsupervised quantitative methods, when combined with close reading and historiographical expertise, function not only as tools of confirmation but also as genuine discovery devices, revealing patterns, continuities, and tensions that remain difficult to grasp through traditional approaches alone. To foster transparency and reuse, we also release an open-source interactive application that allows readers to explore the clusters, indicators, and interpretive pathways underlying our analysis.