Probabilistic Detection of Context-Sensitive Spelling Errors

1
Probabilistic Detection of Context-Sensitive
Spelling Errors

• Johnny Bigert
• Royal Institute of Technology, Sweden
• johnny_at_kth.se

2
What?

• Context-Sensitive Spelling Errors
• ExampleNice whether today.
• All words found in dictionary
• If context is considered,the spelling of whether
  is incorrect

3
Why?

• Why do we need detection of context-sensitive
  spelling errors?
• These errors are quite frequent (reports on
  16-40 of all errors)
• Larger dictionaries result in more errors
  undetected
• They cannot be found by regular spell checkers!

4
Why not?

• What about proposing corrections for the errors?
• An interesting topic,but not the topic of this
  article
• Detection is imperative,correction is an aid

5
Related work?

• Are there no algorithms doing this already?
• A full parser is perfect for the job
• Drawbacks
• high accuracy is required
• not available for many languages
• manual labor is expensive
• not robust

6
Related work?

• Are there no other algorithms?
• Several other algorithms (e.g. Winnow)
• Some do correction
• Drawbacks
• They require a set of easily confused words
• Normally, you dont know your spelling errors
  beforehand

7
Why?

• What are the benefits of this algorithm?
• Find any error
• Avoid extensive manual work
• Robustness

8
How?

• Prerequisites
• We use PoS tag trigram frequenciesfrom an
  annotated corpus
• We are given a sentence, and apply a PoS tagger

9
How?

• Basic assumption
• If any tag trigram frequency is low, that part
  is probably ungrammatical

10
But?

• But dont you often encounter rare or unseen
  trigrams?
• Yes, unfortunately
• We modify the notion of frequency
• Find and use other, syntactically close PoS
  trigrams

11
Close?

• What is the syntactic distance between two PoS
  tags?
• A probability that one tag is replaceable by
  another
• Retain grammaticality
• Distances extracted from corpus
• Unsupervised learning algorithm

12
Then?

• The algorithm
• We have a generalized PoS tag trigtram frequency
• If frequency below threshold, text is probably
  ungrammatical

13
Result?

• Summary so far
• Unsupervised learning
• Automatic algorithm
• Detection of any error
• No manual labor!
• Alas, phrase boundaries cause problems

14
Phrases?

• What about phrases?
• PoS tag trigrams overlapping two phrases are very
  productive
• Rare phrases, rare trigrams
• Transformations!

15
Transform?

• How do we transform a phrase?
• Shallow parser
• Transform phrases to most common form
• Normally, the head
• Benefits retain grammaticality, less rare
  trigrams, longer tagger scope

16
Example?

• Example of phrase transformation
• Only the paintings that are old are for sale
• Only the paintings are for sale

NP
NP
17
Then what?

• How do we use the transformations?
• Apply tagger to transformed sentence
• Run first part of algorithm again
• If any transformation yield only trigrams with
  high frequency,sentence ok
• Otherwise, probable error

18
Result?

• Summary
• Trigram part, fully automatic
• Phrase part, could use machine learning of rules
  for shallow parser
• Finds many difficult error types
• Threshold determines precision/recall trade-off

19
Evaluation?

• Fully automatic evaluation
• Introduce artificial context-sensitive spelling
  errors (using software Missplel)
• Automated evaluation procedure for 1, 2, 5, 10
  and 20 misspelled words(using software AutoEval)

20
Results? 1 errors
21
Results? 2 errors