Owning the agenda: using machine learning to observe the dynamics of issue salience over an election campaign.

Abstract

We analyse a large corpus of 325,000 social media posts from parties, candidates, interest groups and media organisations generated ahead of the 2022 Australian federal election. Learning the topics of these posts provides insight into the issues and campaign narratives of the election. Observed changes in topic prevalence over the campaign – and between different publishers – lets us chart the competition among rival campaign frames and for issue ownership. Measures of user interactions with posts (aggregated to topics) further reveal the dynamics of this competition. We use these data to assess the extent to which parties, candidates and media organisation engineer issue salience, or respond to the public’s appetite for issues and frames revealed in social media interactions. Our analysis of the social media posts goes well beyond the bag-of-words/LDA toolkit firmly established in the analysis of political texts; we represent the posts by embedding their sentences in high-dimensional vector spaces using models trained on massive English language corpora; sentence embeddings preserve word context and hence the semantic distinctiveness of the recovered topics. Further, hierarchical clustering methods help us assess the rich topic space spanned by our corpus in an unsupervised approach, the structure of the topic hierarchy guiding the construction of higher-order topics, which we interpret as “frames” or “issue bundles” in the context of an election campaign.