Appendix I
Word Senses and WordNet
Appendix: Word Senses and WordNet, 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.
- Senses p. 2
- How a sense is defined p. 2
- Why the number of senses is not fixed p. 3
- Relations between senses p. 4
- WordNet p. 5
- The database p. 5
- Sense relations stored there p. 7
- Disambiguation p. 8
- The task and datasets p. 9
- Contextual embeddings p. 10
- Feature-based WSD p. 12
- Lesk as a baseline p. 12
- Word-in-context evaluation p. 13
- Wikipedia as training data p. 14
- Beyond a fixed inventory p. 14
- Using a thesaurus to change embeddings p. 14
- Inducing senses rather than selecting them p. 15
FAQ
Does a dictionary sense inventory define the only possible meanings?
No. Inventories split senses differently. Word-in-context evaluation can compare uses without forcing a shared sense list.
What does the Lesk baseline compare?
Overlap between the context and a sense's definition or example. It is a baseline, not a neural disambiguator.
Is a contextual embedding a sense label?
No. It is a vector that can be used to choose a sense, or to decide that two uses match. The vector itself is not a WordNet synset.
What is sense induction?
Discovering clusters of uses from data instead of selecting from a fixed inventory. The clusters still need an evaluation. They are not automatically dictionary senses.