Researchers To Create AI Tool for Studying Political Conflicts

The frequency, type and geographical location of political conflict events can provide critical information to corporations and governments on where it’s safe to operate. But a large-scale tracking of events like political protests, arms transfers and armed conflict can be a time-consuming, expensive process.

A University of Texas at Dallas research team is working to ease the process by developing a framework that uses machine learning to analyze where and when political conflict events are happening and to predict occurrences.

Dr. Patrick T. Brandt, professor of political science in the School of Economic, Political and Policy Sciences, is studying the computerized extraction of conflict event data from news sources at a global scale using artificial intelligence (AI) and large language models (LLM). The National Science Foundation (NSF) has awarded Brandt and his team a $1.5 million grant (2311142) to develop the framework.

The project builds upon the team’s previous NSF-funded efforts to create ConfliBERT, a publicly available LLM to support conflict studies. Brandt developed ConfliBERT with colleagues and students including Dr. Latifur Khan, professor of computer science in the Erik Jonsson School of Engineering and Computer Science at UT Dallas and a co-principal investigator on the grant.

Analyzing political conflict events has previously been done by humans painstakingly reading thousands of news articles and categorizing events by type, location and other information.

“Doing this by hand costs thousands of dollars, and an LLM costs pennies or dollars. Using machine learning for this work also saves significant time.”

Dr. Patrick T. Brandt, professor of political science

“Doing this by hand costs thousands of dollars, and an LLM costs pennies or dollars,” Brandt said. “Using machine learning for this work also saves significant time.”

Brandt and his team don’t categorize the data themselves; they are developing an algorithm that manages that. “We’re trying to be the tool provider, not the person who builds and maintains the dataset,” he said.

Teaching the LLM to select political stories out of thousands of articles can be tricky.

“For example, we have to filter out all the news stories about sports,” Brandt said. “The language we use to describe sports is a very militarized language; it kind of sounds like a civil conflict.

“The World Cup is always a problem. If you look at all the news reports about the World Cup, you’d think that Argentina, Chile and South America in general had suddenly gone to war with Spain, Germany and Britain.”

After categorizing political conflict events, the next step is for a human to identify the political actors who are engaged in conflict, cooperation and negotiation, Brandt said.

“For example, is an event multiple reports of a single protest, or is it multiple protests?” Brandt said. “Then the human element comes in. A political scientist is able to use the information to identify patterns and determine whether the protests are tied to a country, a rebel movement or some other factor.”

The information gathered by the framework has broad applications, Brandt said.

“One way it’s used is to create a predictive model about political risk,” he said. “That tells a nation, company or industry whether it’s facing geopolitical risk by operating in another country. It can also tell us whether risk has gone up or down in certain regions.”

For example, analyzing tens of thousands of news reports and identifying those about political protests can indicate whether a region is politically stable and worthy of investment, Brandt said. The data also can inform governments about national security decisions.

The researchers’ new efforts will expand the existing ConfliBERT framework to incorporate multilingual settings, including Arabic and Spanish.

“Ultimately, our work can help researchers and policymakers better understand conflict in foreign locations with high accuracy and in real time,” Brandt said.