PHP Repairing Bad Text

Question

This is something I'm working on and I'd like input from the intelligent people here on StackOverflow.

What I'm attempting is a function to repair text based on combining various bad versions of the same text page. Basically this can be used to combine different OCR results into one with greater accuracy than any of them individually.

I start with a dictionary of 600,000 English words, that's pretty much everything including legal and medical terms and common names. I have this already.

Then I have 4 versions of the text sample.

Something like this:

$text[0] = 'Fir5t text sample is thisline';
$text[1] = 'Fir5t text Smplee is this line.';
$text[2] = 'First te*t sample i this l1ne.';
$text[3] = 'F i r st text s ample is this line.';

I attempting to combine the above to get an output which looks like:

$text = 'First text sample is this line.';

Don't tell me it's impossible, because it is certainly not, just very difficult.

I would very much appreciate any ideas anyone has towards this.

Thank you!

My current thoughts:

Just checking the words against the dictionary will not work, since some of the spaces are in the wrong place and occasionally the word will not be in the dictionary.

The major concern is repairing broken spacings, once this is fixed then then the most commonly occurring dictionary word can be chosen if exists, or else the most commonly occurring non-dictionary word.

Answer 1

Have you tried using a longest common subsequence algorithm? These are commonly seen in the "diff" text comparison tools used in source control apps and some text editors. A diff algorithm helps identify changed and unchanged characters in two text samples. http://en.wikipedia.org/wiki/Diff

Some years ago I worked on an OCR app similar to yours. Rather than applying multiple OCR engines to one image, I used one OCR engine to analyze multiple versions of the same image. Each of the processed images was the result of applying different denoising technique to the original image: one technique worked better for low contrast, another technique worked better when the characters were poorly formed. A "voting" scheme that compared OCR results on each image improved the read rate for arbitrary strings of text such as "BQCM10032". Other voting schemes are described in the academic literature for OCR.

On occasion you may need to match a word for which no combination of OCR results will yield all the letters. For example, a middle letter may be missing, as in either "w rd" or "c tch" (likely "word" and "catch"). In this case it can help to access your dictionary with any of three keys: initial letters, middle letters, and final letters (or letter combinations). Each key is associated with a list of words sorted by frequency of occurrence in the language. (I used this sort of multi-key lookup to improve the speed of a crossword generation app; there may well be better methods out there, but this one is easy to implement.)

To save on memory, you could apply the multi-key method only to the first few thousand common words in the language, and then have only one lookup technique for less common words.

There are several online lists of word frequency. http://en.wiktionary.org/wiki/Wiktionary:Frequency_lists

If you want to get fancy, you can also rely on prior frequency of occurrence in the text. For example, if "Byrd" appears multiple times, then it may be the better choice if the OCR engine(s) reports either "bird" or "bard" with a low confidence score. You might load a medical dictionary into memory only if there is a statistically unlikely occurrence of medical terms on the same page--otherwise leave medical terms out of your working dictionary, or at least assign them reasonable likelihoods. "Prosthetics" is a common word; "prostatitis" less so.

If you have experience with image processing techniques such as denoising and morphological operations, you can also try preprocessing the image before passing it to the OCR engine(s). Image processing could also be applied to select areas after your software identifies the words or regions where the OCR engine(s) fared poorly.

Certain letter/letter and letter/numeral substitutions are common. The numeral 0 (zero) can be confused with the letter O, C for O, 8 for B, E for F, P for R, and so on. If a word is found with low confidence, or if there are two common words that could match an incompletely read word, then ad hoc shape-matching rules could help. For example, "bcth" could match either "both" or "bath", but for many fonts (and contexts) "both" is the more likely match since "o" is more similar to "c" in shape. In a long string of words such as a a paragraph from a novel or magazine article, "bath" is a better match than "b8th."

Finally, you could probably write a plugin or script to pass the results into a spellcheck engine that checks for noun-verb agreement and other grammar checks. This may catch a few additional errors. Maybe you could try VBA for Word or whatever other script/app combo is popular these days.

Answer 2

Tackling complex algorithms like this by yourself will probably take longer and be more error prone than using a third party tool - unless you really need to program this yourself, you can check the Yahoo Spelling Suggestion API. They allow 5.000 requests per IP per day, I believe.

Others may offer something similar (I think there's a bing API, too).

UPDATE: Sorry, I just read that they've stopped this service in April 2011. They claim to offer a similar service called "Spelling Suggestion YQL table" now.

Answer 3

This is indeed a rather complicated problem.

When I do wonder how to spell a word, the direct way is to open a dictionary. But what if it is a small complex sentence that I'm trying to spell correctly ? One of my personal trick, which works most of the time, is to call Google. I place my sentence between quotes on Google and count the results. Here is an example : entering "your very smart" on Google gives 13'600k page. Entering "you're very smart" gives 20'000k pages. Then, likely, the correct spelling is "you're very smart". And... indeed it is ;)

Based on this concept, I guess you have samples which, for the most parts, are correctly misspelled (well, maybe not if your develop for a teens gaming site...). Can you try to divide the samples into sub pieces, not going up to the words, and matching these by frequency ? The most frequent piece is the most likely correctly spelled. Prior to this, you can already make a dictionary spellcheck with your 600'000 terms to increase the chance that small spelling mistakes will alredy be corrected. This should increase the frequency of correct sub pieces.

Dividing the sentences in pieces and finding the right "piece-size" is also tricky.

What concerns me a little too : how do you extract the samples and match them together to know the correctly spelled sentence is the same (or very close?). Your question seems to assume you have this, which also seems something very complex for me.

Well, what precedes is just a general tip based on my personal and human experience. Donno if this can help. This is obviously not a real answer and is not meant to be one.

Answer 4

You could try using google n-grams to achieve this.