3 ms·
This is posted here about every three months. Norvig's approach assumes a very small fixed vocabulary (so it won't scale well to the basic multilingual plane),
by kylebgorman 12y ago
This is posted here about every three months.
Norvig's approach assumes a very small fixed vocabulary (so it won't scale well to the basic multilingual plane), and can only recover errors with an edit distance of two or less.
If you are interested in a vastly superior yet still very intuitive approach to spelling correction, I recommend the classic paper on noisy channel model corrector by Kernighan, Church, and Gale, which fuses the probability distribution the distribution over possible errors and the probabilities over words themselves.
http://www.aclweb.org/anthology/C90-2036.pdf http://www.aclweb.org/anthology/C90-2036.pdf
- antics 12y agoI actually think Norvig is making a totally different point than you are. I also think (with respect) that it's a much more important point to learn. Where your point is that there is a better but almost-as-intuitive approach, Norvig's point is basically that a naive application of probability theory is often dramatically easier to implement, and also "good enough" when compared to a (usually much more complicated) rule-based implementation. That's really a great lesson to learn. It's (IMHO) much less valuable to learn that there are better approaches out there -- of course there are.