Text segmentation: dictionary-based word splitting

问题:

Background

Split database column names into equivalent English text to seed a data dictionary. The English dictionary is created from a corpus of corporate documents, wikis, and email. The dictionary (lexicon.csv) is a CSV file with words and probabilities. Thus, the more often someone writes the word "therapist" (in email or on a wiki page) the higher the chance of "therapistname" splits to "therapist name" as opposed to something else. (The lexicon probably won't even include the word rapist.)

Source Code

• TextSegmenter.java @ http://pastebin.com/taXyE03L
• SortableValueMap.java @ http://pastebin.com/v3hRXYan

Data Files

• lexicon.csv - http://pastebin.com/0crECtXY
• columns.txt - http://pastebin.com/EtN9Qesr

Problem (updated 2011-01-03)

When the following problem is encountered:

dependentrelationship::end depend ent dependent relationship
end=0.86
ent=0.001
dependent=0.8
relationship=0.9

These possible solutions exist:

dependentrelationship::dependent relationship
dependentrelationship::dep end ent relationship
dependentrelationship::depend ent relationship

The lexicon contains words with their relative probabilities (based on word frequency): dependent 0.8, end 0.86, relationship 0.9, depend 0.3, and ent 0.001.

Eliminate the solution of dep end ent relationship because dep is not in the lexicon (i.e., 75% word usage), whereas the other two solutions cover 100% of words in the lexicon. Of the remaining solutions, the probability of dependent relationship is 0.72 whereas depend ent relationship is 0.00027. We can therefore select dependent relationship as the correct solution.

Related

• How to separate words in a "sentence" with spaces?
• Top Coder - Text Segmentation Presentation 1/2
• Top Coder - Text Segmentation Presentation 2/2
• Linear Text Segmentation using Dynamic Programming Algorithm
• Dynamic Programming: Segmentation
• Dynamic Programming: A Computational Tool

Question

Given:

// The concatenated phrase or database column (e.g., dependentrelationship).
String concat;

// All words (String) in the lexicon within concat, in left-to-right order; and
// the ranked probability of those words (Double). (E.g., {end, 0.97}
// {dependent, 0.86}, {relationship, 0.95}.)
Map.Entry<String, Double> word;

How would you implement a routine that generates the most likely solution based on lexicon coverage and probabilities? For example:

for( Map.Entry<String, Double> word : words ) {
  result.append( word.getKey() ).append( ' ' );

  // What goes here?

  System.out.printf( "%s=%f\n", word.getKey(), word.getValue() );
}

Thank you!

回答1:

Peter Norvig has written some stuff in python.

http://norvig.com/ngrams/ngrams.py

contains a function called segment. It run a Naive Bayes probability of a sequence of words. works well. Can be a good basis for what your trying to accomplish in java.

If you get it converted to java, i'd be interested in seeing the implementation.

Thanks.

Mike

回答2:

I would approach the problem slightly differently. It's significant that "end" and "dependent" overlap, yet that is lost in your word Map. If instead of a single word map you were to create a set of word maps, with each one representing a possible segmentation of the column name, consisting only of non-overlapping words, you could compute a score for each segmentation based on the word's probabilities and word length. The score for a segmentation would be the average of the scores of the individual words in the segmentation. The score for an individual word would be some function of the word's length(l) and probability(p), something like

score=al + bp

where a and b are weights that you can tweak to get the right mix. Average the scores for each word to get the score for the segmentation and pick the segmentation with the highest score. The score function would not have to be a linear weighting either, you could experiment with logarithmic, exponential or higher-order polynomials (squares for instance)

回答3:

Your problem is a very common one in NLP - don't start by reinventing the wheel - it will take you a long time and not be as good as what is out there already.

You should certainly start by seeing what the NLP libraries have to offer: http://en.wikipedia.org/wiki/Natural_language_processing and http://en.wikipedia.org/wiki/Category:Natural_language_processing_toolkits. Your problem is a common one and there are different approaches which you will need to explore for your corpus.

Your wordsplitting may be found under hyphenation routines. Two possible approaches are n-grams (where the frequency of (say) 4-character substrings are used to predict boundaries) and tries which show common starts or ends to words. Some of these may help with misspellings.

But there is no trivial answer - find what works best for you.