Date:         Fri, 7 Aug 1998 14:09:08 -0700
Reply-To:     "Self, Karsten" <kself@VISA.COM>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         "Self, Karsten" <kself@VISA.COM>
Subject:      Re: FUZZY Matching
Comments: To: "msz03@health.state.ny.us" <msz03@health.state.ny.us>
Content-Type: text/plain

I don't believe numeric to alpha coding then blocking via soundex would
be effective.  Soundex classifies English-language words by phonetic
components.  It doesn't translate well to non-English words, let along
random or arbitrary alpha patterns.
 
Summing digits blocks fairly effectively (values from 0 to 81 with 9
digits), and I've been playing with this.  Once I've blocked values, I
can run a SPEDIS comparison on the original numeric strings -- this
works because SPEDIS is a transform-cost function, and isn't specific to
alphabetic codings, phonetics, linguistic frequency or probability.  The
usual problem of interpreting the SPEDIS score and deciding on a
specific cutoff applies.
 
For those who've suggested going with a commercial solution, the current
problem is that my client *has* chosen a commercial solution (one of the
best), and is trying to evaluate its effectiveness.  Hence the
independent analysis.
 
--
Karsten M. Self (kself@visa.com)
Trilogy Consulting
 
    What part of "Gestalt" don't you understand?
 
> ----------
> From:         msz03@health.state.ny.us[SMTP:msz03@health.state.ny.us]
> Sent:         Friday, August 07, 1998 6:37 AM
> To:   Self, Karsten
> Subject:      Re: FUZZY Matching
>
> Hi.  I haven't done blocking with a SPEDIS-transformed variable.  I
> read
> your note on SSN matching.  It's just a random thought, but I wonder
> if you
> just assigned a consonant to each digit and generated a SOUNDEX how it
> would work out?
>
>