0

I have a small C# project that reads a file and gives me an output: a string that does not contain spaces nor any types of punctuation marks. It may also contain a few misspellings.

Ex. Output:
THEQUICKBROWNFOXJUMPSOVERTHELAZYDOG

I wonder if there is a way to analyze this string by using text mining/data mining and/or regular expressions to indentify the words (preferably nounds, verbs and so forth.) in the string?

I want to read a bunch of files giving me different outputs and put them in statistical order from the one with the most found words to the one which only contain a string of mumbo jumbo.

Also, if the string contain misspellings like:
THEQUICGBROWNFOSJUMPSOVERTHHLAZYDOG
I know that regular expression can calculate the "distans" from a misspelled word and find the most matching one (using a corpus and probability) but this might prove more of a challenge as the string does not have any spaces or punctuation marks. Any ideas how I can solve this?

  • Sharing your research helps everyone. Tell us what you've tried and why it didn’t meet your needs. This demonstrates that you’ve taken the time to try to help yourself, it saves us from reiterating obvious answers, and most of all it helps you get a more specific and relevant answer. Also see How to Ask – gnat Apr 16 '14 at 9:12
  • I will edit the question and show what I have tried so far shortly. I have seen an app with a similar goal like mine here: itunes.apple.com/se/app/beale-decoder/id441926128?mt=8 – RKrogh Apr 16 '14 at 9:21
  • 2
    How would you handle the case where there are multiple possibilities but only one makes sense, e.g., ...OVER THE... vs. ...OVERT HE...? – Blrfl Apr 16 '14 at 12:21
  • 1
    Please don't cross post - stackoverflow.com/questions/23103608/… – ChrisF Apr 16 '14 at 17:55
3

Here is the general approach:

  1. Read a dictionary file and organize all words in a trie data structure. Many Unix systems have such files in the /usr/share/dict/ directory.

  2. Find possible matches of a prefix of your input in the trie. This will usually produce multiple matches, for example theologyisabout begins with theology and the.

  3. If we remove the matched prefixes, we get a set of possible continuations, on which we repeat step 2.

We then end up with a vast tree of possible interpretations.

There are two problems with this:

  • there will be an exponential amount of interpretations
  • we might miss interpretations because of an unknown word, or some unknown grammatical form

We can solve both of these problems by fuzzy matching. When we look up prefixes in the trie, we allow letters to be missing, inserted or changed. However, each such aberration increases the Levenshtein distance. If one interpretation has a too high summed Levenshtein distance, we can prune that interpretation and concentrate on other branches. You could also keep the branches in a priority queue and always investigate the branches with the lowest current edit distance, which is most likely to be a sensible interpretation – not unlike Dijkstra's pathfinding algorithm.

Note that multiple prefix sequences with different edit distances might lead to the same remaining string. You can therefore keep your progress in a data structure that allows parts to be shared. This caching will likely be beneficial for performance. If you in fact try to implement a variant of Dijkstra's algorithm here, a known tail would correspond to a visited node in the graph.

The difficult part is how to actually perform the fuzzy matching. E.g. you could decide on a maximum edit density of x edits per character (0 <= x <= 1), and abort an interpretation if it is guaranteed that this interpretation will have a higher density. For a given string with length l we can therefore determine an edit budget b = x · l. This budget is less important when matching prefixes in the trie, but this trie is only useful if there are less edits than characters in the prefix. An edit budget like b = floor(c / 2) with a prefix of length c might be sensible. How much edits you allow is not only a metric for how garbled texts you allow your system to “understand”, but also a performance setting – smaller budgets run faster, as less alternatives have to be investigated.

  • Wow, thank you very much for a great answer. I will try to implement this and will get back with how it is working. I understand the difficulcy in finding the fuzzy matchings but as I read through your answer I realize it might be enough to extract statistics from finding the words that do exist without a misspelling. I am so very gratefull for the answer. Thank you! – RKrogh Apr 16 '14 at 10:53
0

Thanks to amon I managed to get the algorithm going!

By using this here code a Trie was implemented and filled with the english dictionary (around 23600 words).

By starting reading from each index in the string, feeding it the next char and then the next until the trie does no longer find any more possible solutions (misspelled word or ending of a real one plus the start of the next), judging this outcome and increase the index by 1 words can be found and analyzed.
V
THEQUICKBROWNFOX... Finds THE

_V
THEQUICKBROWNFOX... Finds HE

__V
THEQUICKBROWNFOX... Finds EQ
and so forth.

In this sequence it is possible to check the weigted edit distance between words and find misspellings. However, due to a lack of time this was never fully implemented in my project.
My project has a more advanced approach to this as it is a statistics tool to run sevreal iterations on a given set of texts so feel free to ask if you have more specific questions and I will answer by the best of my capabilities.

Thank you for all the help in this!

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.