Entity linking
Introduction
A research project at BUT FIT that links person mentions across a corpus of OCR'd historical documents. The data has no authority identifiers (Wikidata, VIAF), so the knowledge base of people has to be built from the mentions themselves.
Goal of the project
Build a high-quality train/test dataset of person clusters and measure how far LLM-generated NER annotations can be trusted.
My approach
First I analyse the corpus (3,600 documents, 2.8 GB) and the NER quality, and index it into SQLite. A locally hosted LLM (Ollama / OpenAI-compatible API) then proposes clusters of mentions of the same person, German passages are translated, and a human reviews the proposals in a web annotation tool built with Svelte.
Result
A repeatable pipeline from raw data to a reviewed dataset, documented with architecture decision records (ADRs).
Photos