Spelling Error - Real Words

Real-Word Spelling Errors

for example:

• leaving in about fifteen minuets
• the design an construction of the system
• can they lave him my messages?
• the study was conducted mainly be John Black

25-40% of spelling errors are real words

Solving Real-World Spelling Errors - Processes

for each word in sentence:

• generate candidate set
• the word itself
• all single-letter edits that are English words
• words that are homophones

choose best candidates

• noisy channel model
• task-specific classifier

given a sentence [𝑤₁, …, 𝑤_𝑘] generate a set of candidates for each word 𝑤_𝑖:

• candidate(𝑤₁) = [𝑤₁, 𝑤₁’, 𝑤₁”, …]
• …
• candidate(𝑤_𝑘) = [𝑤_𝑘, 𝑤_𝑘’, 𝑤_𝑘”, …]

choose sequence 𝑊 that maximizes 𝐏(𝑊)

Example Noisy Channel For Real-Word Spell Correction

Simplification: One Error Word Per Sentence

• out of all possible sentence with one word replaced
  • 𝑤₁, 𝑤”₂, 𝑤₃, … two on thew
  • 𝑤₁, 𝑤₂, 𝑤‘₃, … two of threw
  • 𝑤'''₁, 𝑤₂, 𝑤₃, … too of thew
• choose sequence 𝑊 that maximizes 𝐏(𝑊)

Where to Get Probabilities

• language model (e.g. unigram, bigram, n-gram) optionally with stupid backoff
• channel/error model
  • same as for non-word spelling correction
  • plus additional probability for no error 𝐏(𝑤|𝑤)

Peter Norvig’s “threw” Example