Collaborative Research: Extending and Further Developing the Militarized Interstate Confrontation Dataset

The Correlates of War Militarized Interstate Dispute (CoWMID) dataset provides the infrastructure for most studies of the onset and escalation of conflict in the United States and globally. However, the dataset is error-prone and inconsistent across time and region. This project alters this condition by further developing and extending the data of all militarized interstate events by two or more countries, across the globe.

The project collects these data using supervised machine-learning models. First, a common set of search terms for international conflicts are used to identify newspaper articles that contain militarized events between two or more countries. Using state-of-the-art advancements in artificial intelligence (AI) for textual analysis and classification, the project employs already-collected newspaper source information to train a model to identify confirmed militarized events.

The source information spans 1993 to 2014. This infrastructure reduces the monetary and time-associated costs of subsequent data collection, providing near real-time data availability for both policymakers and conflict researchers who investigating the causes and consequences of international conflict.

The project uses a now-common process for identifying and coding large amounts of text. The first part of the project collects and cleans the text corpora (newspaper article coverage of MICs). All newspaper articles will then be tokenized into paragraphs, converting the text into units of analysis.

In the text input phase, each paragraph is attributed a feature if it contains information relevant to a variable with the classification described in the MIE codebook. Finally, these labelled datasets can be used to fine-tune the BERT classifier, building models for each variable to code new events automatically. Since these variable-level classifications are built for the first time, graduate students and faculty jointly hand-code the newspaper articles as well.

Traditional coding helps train the model-building process, and the parallel process serves as a check for accuracy, performance, and quality on both approaches to coding. At the end of the project, variable coding will be completed by the variable models.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.