Mapping Political-Elite Networks in Europe with a Multilingual Joint Entity-Relation Extraction Pipeline

Research Approach

This paper addresses the challenge of mapping informal political-elite networks, which are typically hidden from formal institutional records. The authors introduce a modular, open-weight pipeline designed to transform massive, multilingual news corpora into structured, temporal knowledge graphs. The system architecture is divided into four main stages: chunking, named-entity recognition (NER), a three-stage entity-linking cascade, and relation extraction using a mixture-of-experts model. By anchoring entities to Wikidata identifiers (QIDs) and enforcing a fixed, domain-specific ontology via guided decoding, the pipeline ensures that extracted relationships are consistent, cross-nationally comparable, and grounded in external knowledge.

Pipeline Architecture

The pipeline utilizes GLiNER-X-Large for entity recognition and a sophisticated three-stage linker—combining exact matching, fuzzy string matching, and dense vector search—to resolve entity mentions to Wikidata. The core extraction task is performed by a Qwen3.6-35B-A3B-FP8 model, which is constrained by a 109-entity and 99-relationship type ontology. This approach allows the system to output directed, signed (positive/negative/neutral), and temporal relationships, effectively capturing the nuances of political cooperation and conflict that simple co-occurrence methods miss.

Validation and Utility

The authors validate the pipeline through a gold-standard spot-check of 3,491 relations, reporting high textual correctness. Furthermore, two case studies demonstrate the system's practical utility: it successfully reconstructs the lifecycle of an Austrian political party and maps the complex, polarized network of state-enterprise patronage in Poland. By providing a replicable, open-weight framework, the authors offer a robust tool for empirical computational social science, enabling researchers to study elite networks at a scale previously requiring intensive manual coding.