Who Influences Whom? Analyzing the Interplay of Mainstream and Outsider Parties in Social Media Campaigns

Understanding how political elites adjust their rhetoric in response to rivals is central to mapping informal influence networks, yet most research lacks tools to capture short-term interactions across parties and across countries. This project introduces a multilingual, scalable workflow that leverages recent advances in large language models (LLMs), embedding-based similarity scores, dictionary- and topic-based measures, and NLP pipelines to analyze elite communication--issue emphasis, populist appeals, party ideology--at high temporal resolution. We apply this framework to more than 8,000 YouTube official campaign videos posted by parties in Spain and the United Kingdom in the year preceding each general election from 2015 to 2024. A central innovation is our embedding-centered measurement architecture, which yields harmonized daily indicators of party ideology comparable across countries and election cycles. For each video transcript, we generate dense sentence- and document-level embeddings using transformer models. These embeddings allow us to (1) compute cosine similarities between videos and party manifestos for each election year, (2) position videos within a continuous ideological space, and (3) trace fine-grained rhetorical drift over time. Embedding manifestos alongside videos provides party-specific ideological anchors, enabling precise estimates of convergence and divergence from formal party positions. We extend this embedding-based approach with a richer measurement strategy that foregrounds LLMs and NLP methods. For ideology, we incorporate a prompt-based LLM measure, which provides cross-lingual benchmarks for left–right positioning, as well as Wordscores relative to party manifestos. Beyond ideology, we rely on dictionaries, BERTopic, and supervised transformer classifiers to extract latent signals—such as cultural and economic emphasis, issue salience, and populism—that feed into our party-level daily time series. Together, these components produce complementary video-level scores for ideology, issue salience, and populist style. Aggregated to the party–day level, these scores form high-frequency series that can be plugged directly into dynamic models of party responsiveness. Using these measures, we assess two dimensions of elite responsiveness: how rhetoric changes as elections near, and who influences whom. First, country-specific regressions reveal systematic pre-election repositioning. In Spain, a proportional system, as the election approaches, parties move away from their manifesto anchors, with mainstream actors occasionally shifting toward ideologically compatible coalition partners, and radical parties adopting more extreme positions. In the UK, a majoritarian system, as elections are getting close, all parties move closer to their initial proposals, with mainstream parties ideologically converging toward the center, while radical actors shift further toward the extremes. Second, vector autoregressive (VAR) models uncover directional influence. Across both countries, mainstream parties lead shifts in ideology and issue emphasis, while outsider parties respond either to mainstream cues or to one another—producing short-term feedback loops in attention, tone, and framing. By integrating embeddings, LLMs, and other NLP tools into a reproducible pipeline, this project shows how these methods can identify who influences whom in elite rhetoric and reveal strategic behavior in the days leading up to an election. More broadly, it provides a transferable template for cross-national elite-network research grounded in automated text and video analysis.