Capturing a News Frame–Comparing Machine-Learning Approaches to Frame Analysis with Different Degrees of Supervision

Abstract

The empirical identification of frames drawing on automated text analysis has been discussed intensely with regard to the validity of measurements. Adding to an evolving discussion on automated frame identification, we systematically contrast different machine-learning approaches with a manually coded gold standard to shed light on the implications of using one or the other: (1) topic modeling, (2) keyword-assisted topic modeling (keyATM), and (3) supervised machine learning as three popular and/or promising approaches. Manual coding is based on the Policy Frames codebook, providing an established base that allows future research to dovetail our contribution. Analysing a large dataset of 12 Austrian newspapers’ EU coverage over 11 years (2009–2019), we contribute to addressing the methodological challenges that have emerged for social scientists interested in employing automated tools for frame analysis. While results confirm the superiority of supervised machine-learning, the semi-supervised approach (keyATM) seems unfit for frame analysis, whereas the topic model covers the middle ground. Results are extensively discussed regarding their implications for the validity of approaches.