Improving the selection of news reports for event coding using ensemble classification

Manual coding of political events from news reports is extremely expensive and time-consuming, whereas completely automatic coding has limitations when it comes to the precision and granularity of the data collected. In this paper, we introduce an alternative strategy by establishing a semi-automatic pipeline, where an automatic classification system eliminates irrelevant source material before further coding is done by humans. Our pipeline relies on a high-performance supervised heterogeneous ensemble classifier working on extremely unbalanced training classes. Deployed to the Mass Mobilization on Autocracies database on protest, the system is able to reduce the number of source articles to be human-coded by more than half, while keeping over 90% of the relevant material.