I was thinking of something along the lines of

$soundex{ $name } = soundex($1) 
    if $name =~ m[^(?:O'|Mac|Mc)([A-Z].*$)];
[download]

Which would only work if the list is nicely capitalised, which going by the few BazB posted, it seems that his data might be. The idea was that a search for Connor or Keefe, would also find O'Connor and O'Keefe, which using unassisted soundex wouldn't find.

#! perl -slw
use strict;
use Text::Soundex;


printf "%20s : %s : %s\n",
      $_, m[^(?:O'|Mac|Mc)([A-Z].*$)]  ? soundex($1) : soundex($_), so
+undex($_)
      for qw[
          Connor O'Connor
          Keefe O'Keefe
          MacDonald McDonald Donald Donaldson O'Donnell Donagal O'Dona
+gal
      ];

__END__
P:\test>test
              Connor : C560 : C560
            O'Connor : C560 : O256
            
               Keefe : K100 : K100
             O'Keefe : K100 : O210
             
           MacDonald : D543 : M235
            McDonald : D543 : M235
              Donald : D543 : D543
           Donaldson : D543 : D543
           O'Donnell : D540 : O354
             Donagal : D524 : D524
           O'Donagal : D524 : O352
[download]

The first column of soundex codes are the assisted ones, the second unassisted. You can see what a difference it makes.

That said. It would screw up sound alikes Magee and MacGee, so it may not be such a good idea. Another possibility would be to store two codes for some names, but then you move into needing to normalise the soundex codes into another table. Which wouldn't be a bad thing, but does add complication.

Just stirring you along a bit. 'Mace' and 'Mack' 'Macca' are just about all the McEdge cases.

If you look at the algorithm you will see it is just this:

      ($f) = /^(.)/;
      tr/AEHIOUWYBFPVCGJKQSXZDTLMNR/00000000111122222222334556/;
      ($fc) = /^(.)/;
      s/^$fc+//;
      tr///cs;
      tr/0//d;
      $_ = $f . $_ . '000';
      s/^(.{4}).*/$1/;
[download]

While there is no doubt it gives useful results it is exceptionally simplistic. The problem you are trying to solve is threefold:

1. Simple mis-spelling
2. Pronuncitation Differences
3. Dropped sylables

To a large extent the Knuth algorithm deals reasonably successfully with 1 and 2. Where it falls over is when there is either a mis-spelling of the first letter or a dropped sylable. As you could see in the example I fuzzed the matches by dropping symbols from the code to deal with this, given that there are only 6 numeric codes in use my estimate that it would on average pull 1% of the DB should have read more like 3% if you shorten the codes.

In my picture of a 'better' algorithm I imagine a mapping to perhaps [A-Z]{4} so you could perhaps do linear displacement rather than just dropping.....Oh well back to the real work Bayes filters are the soup de jour.

cheers

tachyon

s&&rsenoyhcatreve&&&s&n.+t&"$'$`$\"$\&"&ee&&y&srve&&d&&print