Researchers present a modular, fully open-weight pipeline for multilingual joint entity-relation extraction that builds signed, temporal knowledge graphs from massive unstructured news corpora. The system combines span-based named-entity recognition with a linking cascade to Wikidata and an ontology-constrained mixture-of-experts model to extract directed relationships.
- The pipeline uses guided decoding to extract directed, signed relationships grounded in a domain ontology.
- A spot-check against a 3491-relation gold standard showed textual correctness ranging from 68.2% strict to 93.7% lenient.
- In Austria, the tool reconstructed a political party's lifecycle, tracking personnel into successor factions and court convictions.
- In Poland, it uncovered overlapping economic and governance networks of state-enterprise patronage and the conflict network between the PO and PiS parties.
By bridging raw multilingual text and structured relational data, this framework provides a robust, replicable foundation for cross-national empirical computational social science.