Entity linking
Paste any text. We detect entity mentions and link each one to a DisamKB entity using LELA, a zero-shot LLM-based entity linker. Hover a highlighted mention for a preview, or click it to open the entity page. Below the text we show the full disambiguation trace — every candidate the linker considered and which one it chose — in the same evidence-first spirit as the rest of the browser.
Linking…
The pipeline, stage by stage.
Linked entities
Solid = linked to a DisamKB entity (hover to preview, click to open). Dashed = detected but no matching entity (NIL).
Disambiguation trace
| Mention | Type | Linked to | Candidates considered (chosen ✓) |
|---|
How it works
The linker is LELA configured against DisamKB. The candidate step swaps LELA's default BM25-only retrieval for DisamKB's observed surface forms (BM25 only fills leftover slots) — the same disambiguation evidence the browser shows humans, now driving the machine.
-
1 Text
Plain-text input — the pasted passage.
-
2 NER
GLiNER · numind/NuNER_Zero-span — zero-shot, 9 labels, threshold 0.5. No fine-tuning.
-
3 Candidate generation
DisamKB surface forms + BM25 fallback (hybrid). Top-20, ranked by observed surface frequency.
-
4 Reranker skipped
LELA offers one; we run NoOp — surface popularity already floats the right entity to the top.
-
5 Disambiguation
gpt-5-mini via OpenAI chat completions. One call per mention, run in parallel; can answer “None” (NIL).
-
6 Entities
Linked DisamKB id, or NIL. Knowledge base (438K entities: id · title · description) powers steps 3 & 5.
Prompt sent to gpt-5-mini
Run a text above to see the real prompt for one of its mentions — system message plus the mention marked in context and the numbered candidate list.
System
User — mention marked in context, then the candidate list ()