[TEXT-131] JaroWinklerDistance: Calculation deviates from definition

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.4
• Fix Version/s: 1.5
• Assignee: Rob Tompkins
• Reporter: Jan Martin Keil
• Created: 02/Aug/18 20:37
• Updated: 18/Oct/18 07:00
• Resolved: 18/Oct/18 07:00

Description

The calculation in JaroWinklerDistance deviates from the definition of the Jaro-Winkler Similarity. By definition the common prefix length is only determine for the first 4 characters. Further, the JaroWinkler is defined as JaroSimilarity + ScalingFactor * CommonPrefixLength * (1 - JaroSimilarity ).
Therefore, I recommend the following changes:

1. Update Jaro-Winkler Similarity calculation

   final double jw = j < 0.7D ? j : j + Math.min(defaultScalingFactor, 1D / mtp[3]) * mtp[2] * (1D - j);
   

   to

   final double jw = j < 0.7D ? j : j + defaultScalingFactor * mtp[2] * (1D - j);
   

1. Update calculation of Common Prefix Length

   for (int mi = 0; mi < min.length(); mi++) {
   

   to

   for (int mi = 0; mi < Math.min(4, min.length()); mi++) {
   

1. Remove unnecessary return value

   return new int[] {matches, transpositions, prefix, max.length()};
   

   to

   return new int[] {matches, transpositions, prefix};