Expanding the utility of protest event datasets with automated variable extraction from event summaries

<p>Previous research shows that civil resistance works. Various event datasets emerged and led to a large body of disaggregated analyses of protest dynamics. Although observations in all event datasets are based on text sources, such as news articles and reports, they provide different and often distinct variables. As a result, scholars with particular research questions and applications are restricted to using one or only a few of these datasets. Many of them are no longer maintained, further restricting applied researchers and other users. However, some datasets provide event summaries for each observation, from which additional variables can be generated. Using the Armed Conflict and Location Event Data, we generate additional variables present in four other datasets with BERT and large language models. Performance varies by target concept. Under their respective primary evaluation settings, supervised encoder models achieve 0.95 average categorical accuracy, while the strongest few-shot decoder model achieves 0.87. Human intercoder reliability similarly varies across categorical variables, with lower agreement concentrated among more interpretive concepts. We enable enhancing the usage of an existing dataset for new descriptive, explanatory, and predictive applications, including future robustness analyses of empirical findings, and ultimately answering previously out of reach research questions due to variable compartmentalization. We contribute to the study of protest and mobilization by showing how natural language processing can be used to reliably expand dataset usage.</p>