limit common prefix in jaro-winkler

[maxbachmann]

The paper for Jaro-Winkler only allows the common prefix to be boosted for the first 4 characters. This can be found in Winklers paper about the Jaro-Winkler similarity (https://files.eric.ed.gov/fulltext/ED325505.pdf). More specifically in

The new string comparator metric basically modifies the basic string comparator according to whether the first few
characters in the strings being compared agree. Specifically, for i=1,2,3,4,

Results still differ from the ones I get in RapidFuzz, since I only boost the score if the Jaro similarity is above 0.7. The paper mentions:

Weight adjustments are only performed for values of cD greater than 0.60. Values below 0.60 are generally associated with pairs of strings associated with nonmatches in U.

This suggests a limit of 0.7. However there is a C implementation from Winkler which only boosts them starting at 0.7 here: https://web.archive.org/web/19990822155334/http://www.census.gov/geo/msb/stand/strcmp.c

I think so far I have only seen implementations which either always boost, or boost starting at 0.7.

[maxbachmann]

This fixes #53

[maxbachmann]

With the second commit it now has the same behavior.

[dguo]

Would it be worth mentioning this in the README and/or the function documentation? Since you said that you've also seen implementations that always boost.

[maxbachmann]

Yes that probably makes sense.

[dguo]

I'll leave it to you to merge this or #67 first.