Which reveals ideology better? Comparing self-presentation and public rhetoric in the Facebook climate debate via embeddings analysis

Information about the ideological orientation of social media users is crucial to analyse a large number of social issues. However, the lack of structured data about users’ beliefs is a common issue in social science. To address this gap, after a review of the state-of-the-art approaches for automated ideology detection, this research focuses on the use of a text-based methodology for this goal, using word embeddings. Specifically, the study contributes to the existing literature by focusing on ideology detection in the specific case of the polarized climate debate, and by testing the less-utilized OpenAI “ada” model for this task. Moreover, this research compares the accuracy of posts, representing users’ public rhetoric, and page descriptions, thought to reveal their self-presentation, to predict users’ ideological orientation. Our findings suggest that post-based methods hold the highest accuracy, but clustering-based approaches using page descriptions also yield respectable results while allowing a reduced use of computational resources. Overall, embedding-based methodology is shown to be a valuable tool for analyzing the ideological leanings of users in polarized debates.