Inferring Political Leaning on X (Twitter): A Zero-Shot Approach in an Italian Scenario

In recent years, there has been growing attention on predicting the political orientation of active social media users, aiding in political forecasts, modeling opinion dynamics, and understanding user polarization. Existing methods, primarily for X (Twitter) users, use content-based or a blend of content, network, and communication analysis. The latest research highlights that a user’s political stance mainly hinges on their views on key political and social issues, prompting a shift towards detecting user stances through their content shared on social networks. This work investigates the use of an unsupervised stance-detection framework Tweets2Stance (T2S) based on zero-shot classification (ZSC) models [1] to predict users’ stances toward a set of social-political statements using content-based analysis of their X (Twitter) timelines in an Italian scenario. The ground-truth user stances are drawn from Voting Advice Applications (VAAs), tools aiding citizens in identifying their political leanings by comparing their preferences with party stances. Leveraging the agreement levels of six parties on 20 statements from VAAs, the study aims to predict Party p’s stance on each statement s using X (Twitter) Party account data. T2S, employing zero-shot learning, proves effective across various contexts beyond politics, showcasing a minimum MAE of 1.13 despite a general maximum F1 value of 0.4, demonstrating significant progress given the task complexity.