Ups... I forgot to say, that the candiate only works if left.length() <= right.length()
!
Ursprüngliche Nachricht
Von: Biedermann,S.,Fa. Post Direkt
Gesendet: Montag, 27. Dezember 2010 12:23
An: 'dev@lucene.apache.org'
Betreff: Improving String Distance calculation performance
Hi,
this is my first post to this mailing list, so I first want to say hello to all of you!
You are doing a great job
In org.apache.lucene.search.FuzzyTermEnum I found an optimised implementation of the LevensteinAlgorithms
which makes use of the fact that the algorithm can be aborted if a given minimum similarity
cannot be reached anymore. I isolated that algorithm into a subclass of org.apache.lucene.spell.StringDistance,
since we usually can make use of this optimisation.
With our current miminum similarity setting of 0.75 this algorithm needs against our test
data only about 22% of run time compared to the original algorithm from the spell package.
With a further optimisation candidate (see below) the runtime can be further reduced by another
third to only 14% of original Levenstein.
So, my first question is: is it worth adding a further method to the StringDistanceInterface:
float getDistance(String left, String right, float minimumSimilarity)
so that applications can make use of possible optimisations (StringDistanceImplementations
without optimisations would just skip the minimSimilarity parameter)?
The idea of the optimsation candidate is about calculating only those fields in the "virtual"
matrix that are near its diagonal.
It is only a candidants since we have not prooven it to work. But with all our test data (0.5
billion comparisons) there is no difference to the original algorithm.
Do you have any counter examples?
Since this is my first post, is this the right mailing list?
Best Regards,
Sven
Here is the code taken from FuzzyTermEnum with some modfications (p and d are initialised
somewhere else):
public float getDistance(final String left, final String right, float minimumSimilarity)
{
final int m = right.length();
final int n = left.length();
final int maxLength = Math.max(m, n);
if (n == 0) {
//we don't have anything to compare. That means if we just add
//the letters for m we get the new word
return (m == 0) ? 1f : 0f;
}
if (m == 0) {
return 0f;
}
// be patient with rounding errors (1.0000001f instead of 1f).
final int maxDistance = (int) ((1.0000001fminimumSimilarity) * maxLength);
if (maxDistance < Math.abs(mn)) {
//just adding the characters of m to n or viceversa results in
//too many edits
//for example "pre" length is 3 and "prefixes" length is 8. We can see that
//given this optimal circumstance, the edit distance cannot be less than 5.
//which is 83 or more precisely Math.abs(38).
//if our maximum edit distance is 4, then we can discard this word
//without looking at it.
return 0.0f;
}
// if no edits are allowed, strings must be equal
if (maxDistance == 0)
return left.equals(right) ? 1f : 0f;
// init matrix d
for (int i = 0; i<=n; i++) {
p[i] = i;
}
// start computing edit distance
for (int j = 1; j<=m; j++) { // iterates through target
int bestPossibleEditDistance = m;
final char t_j = right.charAt(j1); // jth character of t
d[0] = j;
//> here is the optimisation candiates
//only iterate through a maxDistance corridor
final int startAt = Math.max(1, j  maxDistance );
final int finishAt = Math.min(n, maxDistance  1 + j);
for (int i=startAt; i<=finishAt; ++i) { // iterates through text
//
// minimum of cell to the left+1, to the top+1, diagonally left and up +(01)
final char t_i = left.charAt(i1);
if (t_j != t_i) {
d[i] = Math.min(Math.min(d[i1], p[i]), p[i1]) + 1;
} else {
d[i] = Math.min(Math.min(d[i1], p[i]) + 1, p[i1]);
}
bestPossibleEditDistance = Math.min(bestPossibleEditDistance, d[i]);
}
//After calculating row i, the best possible edit distance
//can be found by found by finding the smallest value in a given column.
//If the bestPossibleEditDistance is greater than the max distance, abort.
if (j > maxDistance && bestPossibleEditDistance > maxDistance) {
//equal is okay, but not greater
//the closest the target can be to the text is just too far away.
//this target is leaving the party early.
return 0.0f;
}
// copy current distance counts to 'previous row' distance counts: swap p and d
int _d[] = p;
p = d;
d = _d;
}
// our last action in the above loop was to switch d and p, so p now
// actually has the most recent cost counts
// this will return less than 0.0 when the edit distance is
// greater than the number of characters in the shorter word.
// but this was the formula that was previously used in FuzzyTermEnum,
// so it has not been changed (even though minimumSimilarity must be
// greater than 0.0)
return 1.0f  ((float)p[n] / (float) (maxLength));
}

To unsubscribe, email: devunsubscribe@lucene.apache.org
For additional commands, email: devhelp@lucene.apache.org
