- Genocides and mass killings are preceded by triggering events
- Examining media reports with machine learning can help identify triggering events
- A virtual presentation at PEARC20 brings this work to a larger audience
Between April 7 and July 15, 1994, violent extremists murdered around 800,000 people in Rwanda. Most of these victims were members of the Tutsi minority ethnic group, and experts estimate that Hutu radicals killed around 75% of the country’s Tutsi population in these 100 days of terror.
Sadly, Rwanda’s situation isn’t unique. Myanmar soldiers recently gave us a glimpse into the August 2017 orders to massacre entire villages of Rohingya Muslims. A recent report found that the treatment of Uighurs in the Xinjiang region of China meets the UN definition of a genocide. And these are just some of the well-known mass killings—there are many others that don’t reach the same level of notoriety.
Timothy Burley, an undergraduate student at University of Notre Dame, is interested to learn how the state can prepare someone to kill. More specifically, he and his colleagues are studying why mass killings happen when they do.
“The question I wanted to ask myself was, ‘Why would humans be treated like this?’” says Burley. “And even more specifically, ‘What brings a human to want to kill lots of people?’”
Specifically, Burley and his colleagues looked to a machine-learning analysis of news articles to better understand the timing and types of events that precede a mass killing. In doing so, they hope to learn more about how these triggers can lead to a government’s decision to kill its own citizens.
Too much for a human to handle
Burley worked with Dr. Ernesto Verdeja of the Kroc Institute for International Peace Studies and Dr. Paul Brenner of the Center for Research Computing to survey a variety of news articles that preceded mass killings and attempt to identify triggers for the violence.
“We retrieved tens of thousands of articles,” says Burley. “The idea was to get all of these documents, clean them, classify them, binarize them, and then feed them into a program called Event Coincidence Analysis that would statistically define them as triggers.”
Previous attempts to create a working dictionary of triggers have been hampered by the inefficiency of human coders. That’s why Burley turned to natural language processing (NLP) to sort through the vast amount of media information.
“For example, one of our triggering events is ‘protest,’” says Burley. “The verb ‘protest’ would be the main word. And then right beneath that you would write all of the context words that you’d think would appear around it. We found that method to be relatively tedious.”
“Relatively tedious” is an understatement. Burley and his colleagues had to sort through more articles than a human could handle. That’s why the group leveraged an NLP machine coding system to flag potential triggers.
Human coders then did a second pass to verify the machine-coded flags. By examining previous work, the researchers identified nine potential triggers for state-led violence:
- Escalations of armed conflict
- Armed conflict spillover from a neighboring country
- Increases in opposition organization/mobilization
- Anti-state riots/demonstrations
- Changes in political control/election loss
- Coups/attempted coups
- Canceled/postponed/contested forthcoming elections
- Changes in peace and negotiation dynamics
Integrating the highly structured characteristics of NLP with the fluid nature of political science is a tricky business. Thankfully, Burley had help from Dr. Verdeja and political science PhD student Angela Chesler. “They're the brains behind all of the political events that we're looking for,” says Burley.
So how does an undergraduate find himself researching some of the worst events to ever transpire in human history?
“My involvement with the entire project started with the Kellogg International Scholars Program,” says Burley. “It was originally me with Dr. Verdeja and a couple of other people on that team. We were using a manual approach to find all the triggering events.”
After achieving some preliminary results, Verdeja presented the group’s findings to a think tank. The team got feedback to expand their search, but doing so would go beyond the abilities of the human coders. To help with this, Burley found the tools necessary to create a machine-learning approach.
This eventually progressed into a fully fledged study that Burley presented at the PEARC20 computing conference. Though the coronavirus meant that the conference took place online, it was still an important experience for Burley.
“I was excited to present because it was my first ACM paper,” says Burley. “It felt like a significant accomplishment to me given that I had no programming experience before college.”
Burley gained that experience thanks to the Research Experience for Undergraduates program at the Notre Dame Center for Research Computing, which gives students a chance to work collaboratively with expert mentors on computational social science projects. The program hopes to develop multidisciplinary social scientists with the computational expertise to analyze our digital traces and answer important social science questions.
How and why societies are tipped into committing mass atrocities is one of those important questions. Looking closely at how governments drive their citizens to kill is a vital step towards predicting and preventing the next genocide before it takes place.
- Mining the news for data
- What presidential speech reveals
- Why do we fight?
- Can AI and video surveillance prevent mob violence?