EDCA: a Vietnamese education knowledge graph and a typed entity-schema layer for WebNLG, REBEL and Wiki-NRE

Supporting data for the paper 'A Canonicalization-Signal Design Space and Assessment Methodology for Schema-Free LLM-Based Knowledge Graph Construction' (Knowledge-Based Systems). Two artifacts. (1) A Vietnamese higher-education knowledge graph built from 10,181 text chunks with no predefined schema: 33,498 typed entities together with the relation and entity-type schemas the pipeline discovered for them. 21.6% of entities carry the fallback type 'Thing', and a residue of code-mixing survives into type and surface names; both are released unrepaired and documented, since an artifact whose defects are cleaned out of the release but not out of the method misrepresents the method. (2) A typed entity-schema layer for the WebNLG, REBEL and Wiki-NRE relation-extraction benchmarks, which ship gold relations but no entity types: every relation slot is assigned the entity type it takes, by LLM inference, then grounding against Wikidata P2302 and DBpedia domain/range constraints, then human adjudication to zero remaining mismatches, with the human overrides recorded separately from the machine grounding. A second, model-independent gold derived from Wikidata instance-of (P31) assertions is included alongside it.