Most algorithms for splitting text into sentences which I've found rely on punctuation being correct. However, in many real world applications, there will be substantial numbers of punctuation errors (missing periods, extraneous periods, etc.) Are there sentence-splitting algorithms which deal with this?
Here's how such a sentence boundary algorithm could be built.
The crucial thing is to come up with sequences of words (n-grams) that have a high likelihood of occurring at the end of a sentence or at the start of a sentence and a small likelihood of occurring within a sentence.
A list of these n-grams could be compiled with a large corpus such as the written part of the British National Corpus. Once the list is complete, the algorithm would go through the text without punctuation and for every word boundary look up the preceding n-grams in the list. For example, in a sequence of words such as
have you seen it it's incredible
once the algorithm is at have you seen it it would look up seen it in the list and give you a score such as: Occurred 56 times at the end of sentences and 22 times within a sentence, i.e. a score of 56/22 = 2.5. There should also be a list of three-word sequences, so look up you seen it, which might have a score of 13/2=6.5, then go on with a four-word list. Now average all scores and if it passes a certain threshold (which would need to be determined empirically) the algorithm sets a sentence boundary.
The same could be done with n-grams/word sequences occurring sentence-initially, and n-grams which occur at sentence boundaries.
Edit: As @P Elliott pointed out in the comments, a somewhat similar algorithm has been implemented and has high accuracy.
|show 5 more comments|