Question

I’m am trying to use a levenshtein algorithm I found on the ‘net to calculate the closest value to a search term. In order to implement fuzzy term matching. My current query runs about 45 seconds long. I’m hoping I can optimize it. I’ve already added indexes for the fields that I’m calculated the levenshtein value for. The levenshtein function I found may not be the most optimized and I take no credit in it’s implementation. Here is that function:

CREATE FUNCTION [dbo].[LEVENSHTEIN]( @s NVARCHAR(MAX), @t NVARCHAR(MAX) )
/*
Levenshtein Distance Algorithm: TSQL Implementation
by Joseph Gama

http://www.merriampark.com/ldtsql.htm

Returns the Levenshtein Distance between strings s1 and s2.
Original developer: Michael Gilleland http://www.merriampark.com/ld.htm
Translated to TSQL by Joseph Gama

Fixed by Herbert Oppolzer / devio
as described in http://devio.wordpress.com/2010/09/07/calculating-levenshtein-distance-in-tsql
*/
RETURNS INT AS
BEGIN

  DECLARE @d NVARCHAR(MAX), @LD INT, @m INT, @n INT, @i INT, @j INT,
    @s_i NCHAR(1), @t_j NCHAR(1),@cost INT

  --Step 1
  SET @n = LEN(@s)
  SET @m = LEN(@t)
  SET @d = REPLICATE(NCHAR(0),(@n+1)*(@m+1))
  IF @n = 0
  BEGIN
    SET @LD = @m
   GOTO done
  END
  IF @m = 0
  BEGIN
    SET @LD = @n
    GOTO done
  END

  --Step 2
  SET @i = 0
  WHILE @i <= @n BEGIN
    SET @d = STUFF(@d,@i+1,1,NCHAR(@i))        --d(i, 0) = i
    SET @i = @i+1
  END

  SET @i = 0
  WHILE @i <= @m BEGIN
    SET @d = STUFF(@d,@i*(@n+1)+1,1,NCHAR(@i))    --d(0, j) = j
    SET @i = @i+1
  END

  --Step 3
  SET @i = 1
  WHILE @i <= @n BEGIN
    SET @s_i = SUBSTRING(@s,@i,1)

    --Step 4
    SET @j = 1
    WHILE @j <= @m BEGIN
      SET @t_j = SUBSTRING(@t,@j,1)
      --Step 5
      IF @s_i = @t_j
        SET @cost = 0
      ELSE
        SET @cost = 1
      --Step 6
      SET @d = STUFF(@d,@j*(@n+1)+@i+1,1,
        NCHAR(dbo.MIN3(
          UNICODE(SUBSTRING(@d,@j*(@n+1)+@i-1+1,1))+1,
          UNICODE(SUBSTRING(@d,(@j-1)*(@n+1)+@i+1,1))+1,
          UNICODE(SUBSTRING(@d,(@j-1)*(@n+1)+@i-1+1,1))+@cost)
        ))
      SET @j = @j+1
    END
    SET @i = @i+1
  END      

  --Step 7
  SET @LD = UNICODE(SUBSTRING(@d,@n*(@m+1)+@m+1,1))

done:
  RETURN @LD
END

And here is the query I’m using:

SELECT [Address], [dbo].[LEVENSHTEIN](@searchTerm, [Address]) As LevenshteinDistance
FROM Streets
Order By LevenshteinDistance

I’m not a DBA so please forgive my ignorance in any best practices – that’s why I’m here to learn :). I really don’t want to offload this processing in the business layer and am hoping to keep it in the data layer but with only 16k records taking 45 seconds to process it’s currently not usable. This is only with a small subset of the records which will comprise the entire data store once I’m done processing the input files. Thanks in advance.

Answer 1

If you want it to run really fast, consider creating a dll in C#. It will improve your speed by 150% 😉

Here is my blog that explains you how to do it.

http://levenshtein.blogspot.com/2011/04/how-it-is-done.html