Entity linking
Paste any text — each mention is linked to its GPTKB v2 entity, disambiguated by context (the same word can resolve to different entities).
Pipeline
0.0sThe linker is LELA (Haffoudhi et al., IJCAI-ECAI 2026), configured against GPTKB v2 — candidate retrieval runs on the KB's observed surface forms, the same disambiguation evidence the browser shows humans.
-
Detect mentions
GLiNER (numind/NuNER_Zero-span) — zero-shot NER, 9 labels, threshold 0.5, no fine-tuning.
-
Gather candidates
GPTKB v2 observed surface forms ranked by frequency; BM25 fills leftover slots (top-20). LELA's optional Reranker stage is deliberately skipped (NoOp) — surface popularity already floats the right entity to the top.
-
Disambiguate
One openai/gpt-oss-120b call per mention, run in parallel, choosing from the numbered candidate list — the model may answer “None” → the mention stays NIL.
Run a text above to see the exact prompt sent for one of its mentions.
Prompt for · model answered
Linked text
No entity mentions detected in this text.
linked opens the entity page (hover for a preview) · NIL no GPTKB v2 match (click to see the rejected candidates)
Disambiguation evidence
Every candidate the linker weighed, per mention, grouped by how it was retrieved — GPTKB v2 surface forms (ranked by observed frequency) and, where those run out, a BM25 text-relevance fallback. The two use different score scales, so each group's bars are normalized on their own.