Christopher Lucas - Transformer-Based Classification of Political Advertisements

Video is an increasingly ubiquitous medium for political communication. However, existing research on the content of political video relies on either manually classifying every video of interest or automated classification with text transcriptions, which can be costly to produce and unreliable in some cases. We propose an alternative method using a transformer model and apply it to the automated classification of negativity in political advertisements on television. This approach uses only the one-dimensional audio signal as input but nonetheless outperforms existing approaches that rely on text and audio features. We then demonstrate that the model successfully extrapolates from television to online advertisements and use it to study how negative advertising differs across these domains. In this application, we also offer guidance on how to robustly use machine learning predictions as proxy variables in statistical tests.