Is a single model enough? The systematic comparison of computational approaches for detecting populist radical right content

Abstract

The rise of populist radical right (PRR) ideas stresses the importance of understanding how individuals engage with PRR content online. However, this task is complicated by the variety of channels through which such engagement can take place. In this article, we systematically compare computational approaches for detecting PRR content in textual data. Using 66 dictionary, classic supervised machine learning, and deep learning (DL) models, we compare how these distinct approaches perform on the PRR detection task for three Germanophone test datasets and how their performance is affected by different modes of text preprocessing. In addition to individual models, we examine the performance of 330 ensemble models combining the above-mentioned approaches for the dataset with a particularly high volume of noise. Our findings demonstrate that the DL models, in combination with more computationally intense forms of preprocessing, show the best performance among the individual models, but it remains suboptimal in the case of more noisy datasets. While the use of ensemble models shows some improvement for specific modes of preprocessing, overall, it mostly remains on par with individual DL models, thus stressing the challenging nature of computational detection of PRR content.