VALPOP article in ArXiv, 2026
Whether a country’s political elites organise into closed coalitions that capture public resources, or into open civic networks that sustain good governance, is one of the oldest questions in political science. It is also one of the hardest to answer with data. The relationships that matter most – informal alliances, business ties, patronage appointments, and rivalries – rarely appear in official registers. They appear in the news.
A new study by VALPOP researchers shows how those relationships can now be extracted from news archives at a scale no human coding team could match.
The problem: elite networks are visible but unmeasured
Existing databases of political elites record formal positions: who was president, who sat in which cabinet. What they omit is the connective tissue of political life, the friendships, board seats, court cases, and conflicts that determine how power actually flows. Assembling that relational picture has traditionally required teams of coders reading biographies article by article, which confines systematic research to a handful of well-funded projects. Earlier automated methods could only count how often two names appeared together, which tells us that two actors share a stage, but not whether they are allies, rivals, employer and employee, or defendant and judge.
The solution: an open, multilingual extraction pipeline
The study by Kirill Solovev and Jana Lasser from the University of Graz introduces the multilingual extraction pipeline developed by VALPOP. Importantly, VALPOP builds a fully open pipeline, meaning every component runs on freely available models rather than proprietary commercial services, that reads news articles in multiple languages and converts them into a knowledge graph: a structured map in which people, parties, firms, and institutions are nodes, and their relationships are labelled connections. The system recognises named entities in the text, matches each one to a unique identifier in Wikidata (a free, language-independent knowledge base), and then uses a large language model to classify the relationship between them, choosing from a fixed vocabulary of 99 relationship types. Crucially, each relationship carries a sign – positive, neutral, or negative – and a timestamp. The result is not a static sociogram but a signed, temporal network: it distinguishes support from criticism and records when ties begin and end.
Solovev, K., & Lasser, J. (2026). Mapping political-elite networks in Europe with a multilingual joint entity-relation extraction pipeline. arXiv Working Paper 2606.27347. https://arxiv.org/abs/2606.27347
Validated in two countries, covering half a million articles each
The study validates the pipeline against events whose facts are fixed in the public record. In Austria, the system processed half a million press articles and reconstructed, without any prior instruction, the complete lifecycle of the BZÖ party: its 2005 secession from the FPÖ, the death of founder Jörg Haider in 2008, the regional breakaway triggered by the Hypo Alpe Adria banking collapse, dated to the correct calendar days, and the scattering of its personnel into successor parties. The network even clustered, unsupervised, the exact group of Carinthian politicians later convicted in the Austrian courts, and showed that the cohort carried nearly four times the legal-jeopardy edge load of a matched control group of politicians.
In Poland, the pipeline mapped the elite system across a quarter century of news. It recovered the well-documented duopoly between Civic Platform (PO) and Law and Justice (PiS), and showed that the two camps’ mutual ties are 95–98% negative, a textbook signature of polarisation that standard network methods cannot detect because they ignore whether ties are friendly or hostile. It also confirmed a central concern of governance research: the firms most densely tied to politicians are overwhelmingly state-owned enterprises, and their leadership turns over sharply after each change of government, with the largest churn following the 2023 election.
Developed within VALPOP for cross-national analysis
This pipeline was developed within VALPOP and is directly relevant to VALPOP’s objectives. One potential application is to distinguish rent-seeking coalitions that extract public resources (O-groups, after economist Mancur Olson) from civic networks that support public-goods provision (P-groups, after Robert Putnam), as proposed by our sister project NET-ROL. This allows researchers to ask how these structures shape populism, rule-of-law erosion, and the value citizens place on public goods. Answering that question comparatively requires observing elite networks across many countries and languages on equal terms. By turning ordinary news coverage into comparable, signed, time-stamped network data and by openly releasing the ontology and gold standard, this study provides the empirical foundation on which VALPOP’s cross-national analysis will be built.
