in reply to Re^3: Fingerprinting text documents for approximate comparison
in thread Fingerprinting text documents for approximate comparison
no offense intended, BrowserUK!)
None taken :)
My statement "no easy way" derives from several attempts that I have had at doing this. I've achieved a degree of success, but it is very easily thrown.
The problem with reducing your signature to a numerical value is that two identical pieces of text, except one has a spelling error in an otherwise insignificant word, promotes that word to the realms of a 'rare' word. When you now derive your numeric value, that typo can completely change the value derived and so, the two pieces of text no longer compare favourably.
My best attempt at working around that was to effectively spell check the document. That is to say, I only used words that appeared in my dictionary file when building my lists of N-rare words.
Whilst that removes typos of common words from being considered rare words, it also excludes many proper nouns and similar that would otherwise form very good components of a signature.
For example, people's names and place names often form a very strong correlative indicator for news stories, but if you only accept "correctly spelt words" as candidates, then you exclude many names that will likely not appear in your dictionary.
|
|---|
| Replies are listed 'Best First'. | |
|---|---|
|
Re^5: Fingerprinting text documents for approximate comparison
by ww (Archbishop) on Mar 24, 2005 at 20:19 UTC |