New method, different war? : evaluating supervised machine learning by coding armed conflict

摘要

The internet promises ad hoc availability of any kind of information. Conflict researchersseem to be bound only by the effort needed to find and extract the necessary informationfrom international news sources. This begs the question of whether the sheer number ofaccessible news sources and the speed of the news cycle dictate an automated coding approachin order to keep up. Will the initial costs of implementing such a system outweighthe possible loss of information on violent conflict? We answer these questions in relationto the Event Data on Armed Conflict and Security project (EDACS) where we carry out bothhuman and machine-assisted coding to generate spatiotemporal conflict event data. We usespatiotemporal comparability measures for quantitative and qualitative comparison of thetwo datasets. While the quality of human-coding exceeds a purely automated approach, acompromise between efficiency and quality results in a supervised, semi-automated machinelearning approach. We conclude by critically reflecting on the possible discrepancies in theanalysis of these resulting datasets.