Chapter 24
Coreference Resolution and Entity Linking
Chapter: Coreference Resolution and Entity Linking, in Daniel Jurafsky and James H. Martin, Speech and Language Processing, 3rd edition draft, August 19, 2026.
Topics to learn
Page numbers refer to the linked chapter PDF.
- Linguistic background p. 4
- Types of referring expressions p. 4
- Information status p. 5
- Non-referring expressions p. 6
- Properties of the coreference relation p. 7
- The task p. 9
- Datasets p. 9
- Mention detection p. 10
- Coreference architectures p. 13
- Mention-pair classifiers p. 13
- Mention ranking p. 14
- Entity-based models p. 15
- Hand-built features p. 15
- Neural span representations, mention scores, and antecedent scores p. 17
- Entity linking p. 20
- Dictionaries and web-graph links p. 20
- Neural graph-based linking p. 22
- Evaluation and failure p. 24
- Coreference scores p. 24
- Winograd schemas p. 25
- Gender bias p. 26
FAQ
Does every noun phrase need a coreference link?
No. Some expressions do not refer, and some mentions are the first introduction of an entity. Mention detection and antecedent selection are separate decisions.
How is mention ranking different from mention-pair classification?
A pair classifier scores one candidate link at a time. A ranker compares antecedent candidates for the same mention, including the choice to start a new entity. Those training signals are not the same.
Is entity linking the same task as coreference?
No. Coreference clusters mentions in a document. Entity linking maps a mention to an entry in an external inventory. A document cluster can be right while the inventory link is wrong.
Why does the chapter treat gender bias as an evaluation issue?
Coreference decisions can depend on stereotypes rather than the text. A score on a standard test set can hide systematically worse or skewed errors on gender contrasts.