Training Strategies for Promoting Accuracy in Crowdsourced Content Analysis

Abstract

In this work, we evaluate different instruction strategies to improve the quality of crowdcoding for the concept of civility. We test the effectiveness of training, codebooks, and their combination through 2x2 experiments conducted on two different populations—students and Amazon Mechanical Turk workers. In addition, we perform simulations to evaluate the trade-off between cost and performance associated with different instructional strategies and the number of human coders. We find that training improves crowdcoding quality, while codebooks do not. We further show that relying on several human coders and applying majority rule to their assessments significantly improves performance.