Curate LabsCurate Labs

Information Extraction

Research into turning documents into facts, links, and evidence.

Information extraction research turns the text that comes out of documents into facts, links, and evidence that a graph can hold and a reviewer can check. The goal is not just to find entities; it is to preserve the relationship, the source span, and the confidence needed to act on them.

Core questions

  • Which entities and relationships matter for the question being asked, and how should a schema or ontology emerge from the documents rather than be imposed on them?
  • How should extraction expose confidence, ambiguity, and missing evidence, so that a graph records what is known and what is only suspected?
  • When should extraction use schemas, graphs, language models, rules, or human review, and how do you measure the choice on a specific document set?

Where it is applied

This research shapes the path from Verdant, which produces structured text, to GraphForge, which holds the entities, relationships, and evidence that come out of it. Earlier work in this thread includes heritage extraction to CIDOC-CRM and text-to-graph experiments, published under DecisionNerd on GitHub.