Early warning of incidents during protest demonstrations using Twitter

Abstract

While most protest demonstrations remain peaceful, some of them become violent, resulting in possibly dangerous situations. Because there is a positive correlation between social media use and protest participation, we propose to analyze incidents during protest demonstrations using social media data. In this study, a Twitter dataset is collected related to a Dutch demonstration where protesters did not comply with COVID-19 rules of the government. Following, an exploratory data analysis is performed to identify the phases of Twitter coverage after an incident during a protest demonstration. Additionally, machine learning models are trained to distinguish incident-related from non incident-related tweets. Furthermore, analysts at the Dutch national police force are interviewed to identify the information need when automatically detecting incidents during protest demonstrations. Lastly, an early warning system is created that automatically extracts tweets and detects incidents during protest demonstrations. Findings show four phases of Twitter coverage can be identified after an incident during a protest demonstration, Support Vector Machines (SVM) perform best in distinguishing incident-related from non incidentrelated tweets and analysts at the Dutch national police force want to obtain incident information as soon as possible. The developed system was able to detect incidents during a protest demonstration by using Twitter data, but could be improved.