Capturing a Frame in Text - Exploring the Employability of Machine-Learning Approaches for Automated Frames Analysis

The increasing availability of large amounts of digitalised text sources calls for an efficient and reliable way of identifying and analysing concepts. Originating in psychology, the concept of framing has become a popular area of research across diverse disciplines, first, in the 1970-s, in sociology, later in communication, and most recently in political science. Focusing on ‘frames in communication’, a sociological understanding of a frame as a way to present the same information on events or issues in different ways has become increasingly popular in communication and political science. When it comes to the empirical identification of frames, scholarship distinguishes between inductive and deductive approaches: Following the inductive, bottom-up approach, frames emerge from the data; the deductive approach, in contrast, investigates the use of so-called generic frames that are defined prior to the analysis and generally applicable across topics. For an automated analysis of text, the identification of frames poses problems for both approaches, with the differentiation between ‘issues’ and ‘frames’ being the most important but not only challenge. Against this broader background, we pose the question under what theoretical and methodological conditions can we automate frames analysis? Despite some recent advancement in automated frame analysis, existing tools suitable to automatically analyse frames in textual data have, to the best of our knowledge, not been systematically contrasted to date. Focusing on both generic and issue-specific frames, we here discuss automated content analysis methods requiring different degrees of supervision. Concretely, we look into (1) topic modelling as an unsupervised machine learning approach, (2) latent semantic scaling as a semi-supervised approach, and (3) supervised machine learning. Our paper contributes to a better understanding of tools existing in communication and political science. It also provides a systematic assessment of the efficiency of these different approaches in terms of resources and human input. In addition, connecting different approaches to the conceptual discussion allows evaluating and selecting the adequate tools for valid measurement of frames using both, deductive and inductive approaches. Overall, our paper provides a methodological contribution to the developing field of automated text analysis in the social sciences.