From TEI Edition to Knowledge Graph: Semantic Enrichment of Early Modern Sermon Corpora with LLM-Based Entity Extraction

Abstract

Historical sermon corpora are typically accessed through document-centric TEI encoding, which limits cross-document discovery and semantic querying. We present HERALD (Homiletical Entity Recognition and Linked Data), a TEI-to-knowledge-graph pipeline demonstrated on the sermon corpus of Skarga, spanning five sermon types. It transforms TEI metadata into RDF, then enriches it with a hybrid approach: rule-based extraction for biblical citations and an LLM-based agentic method for Named Entity Recognition and Disambiguation (NERD) of persons and places, both governed by an OWL 2 DL ontology aligned to CIDOC-CRM, LRMoo, and PROV-O. The resulting graph, loaded into Neo4j, supports intertextual and cross-type queries unavailable in document-centric processing. Evaluation on a 20 sermon manually annotated gold standard and cross-validation on modern Polish sermons confirm extraction quality and generalisability. The pipeline is designed as a reusable enrichment layer for TEI-encoded digital library collections.