Appendix D
Spelling Correction and the Noisy Channel
Appendix: Spelling Correction and the Noisy Channel, 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.
- The noisy channel p. 2
- A language model for the intended word p. 2
- A channel model for how the error was produced p. 2
- Error types p. 7
- Non-word errors p. 2
- Real-word errors p. 7
- Better channel models p. 10
- Partitions of edit operations p. 10
- Pronunciation p. 10
FAQ
What are the two probabilities in the noisy channel?
One is how likely the candidate was the intended word. The other is how likely the observed error is given that candidate. Correction ranks their product, not edit distance alone.
Why can a real-word error survive a dictionary check?
The observed string is a word, just not the intended one. A non-word filter never fires. The channel has to consider confusable real words.
Does a shorter edit distance always win?
No. A frequent intended word, or a more probable error type, can outrank a rarer word with a smaller edit. The model defines the ranking.
What does a pronunciation model add?
Some errors are attempts to spell a sound, not arbitrary character edits. A channel that only counts insert, delete, and substitute misses that route.