LLM‑Extracted Political Networks as a Basis for Ideological Mapping: Evidence from Chilean Elites, 2019–2025

# LLM‑Extracted Political Networks as a Basis for Ideological Mapping: Evidence from Chilean Elites, 2019–2025 Naim Bro, Benjamín Palacios, Andrés Abeliuk, Martín Ríos This paper proposes a exploratory contribution to debates on how LLM‑assisted extraction can broaden the kinds of political network questions we can ask and answer. It introduces a flexible methodology for mapping ideological positions from signed political networks derived automatically from news text. Building on recent advances in LLM-assisted text-to-graph extraction (Bro 2025) and computational models of antagonism, alignment, and cleavage structure in signed networks (Fraxanet et al. 2024), we propose a framework for recovering latent ideological structure that capture how Chilean elites repositioned themselves during the 2019–2025 period. The approach is demonstrated through a critical case: the 100 historically center-left figures who publicly endorsed a center-right candidate during the presidential election of 2025, Evelyn Matthei (La Tercera, 2025). The core premise is that positive ties reflect ideological affinity and negative ties capture ideological distance (Bro 2025), such that the global structure of a signed network encodes an underlying ideological space. Using a corpus of 60,000 media articles mentioning Chilean elites, we extract a dynamic signed adjacency matrix at six‑month intervals. Because a central theme of the workshop concerns reliability, replicability, and the comparative value of different extraction and modeling strategies, we explicitly refrain from committing to a single technique. Instead, we outline a modular framework that can incorporate multiple latent‑space approaches—including embeddings, signed‑graph models, and frustration‑minimizing methods—to recover actors’ evolving positions over time, and to compare how methodological choices affect substantive conclusions. Substantively, we test the hypothesis that Chile’s long-dominant dictatorship/democracy cleavage eroded after 2019, giving rise to a new axis structured around order, security, and technocratic governance. The Matthei-signatory group constitutes an ideal probe: their network-derived locations in 2019 cluster firmly within the traditional center-left, but their trajectories show a gradual drift toward the emerging center-right pole. By 2025, their positions converge with right-of-center elites, consistent with the political realignment implied in their endorsement letter. These temporal shifts allow us to quantify not only ideological movement but also the changing geometry of Chile's elite cleavage structure. Methodologically, the project examines how different extraction and modeling strategies shape the resulting ideological space. Rather than presenting a fixed pipeline, it demonstrates how network-based ideal points can serve as a generalizable alternative to roll-call-based methods such as NOMINATE and text-based ideal points, while also showing how different signed-network models may produce convergent or divergent ideological structures. Because networks reflect relationships rather than votes or word use, this approach extends ideal point estimation to contexts lacking legislative data, to multi-elite systems, and to moments of rapid realignment when traditional signals are weak.