in reply to Re: Re: tokenize plain text messages
in thread tokenize plain text messages

Try this version of your ....oneregex sub. (Phew! Long names:). 

D:\Perl\test>257026
192.168.1.1 | $2.50 | Keep | example | and | is | This | 1,500 | an
          Rate   lists strings   regex
lists    757/s      --    -56%    -61%
strings 1736/s    129%      --    -10%
regex   1930/s    155%     11%      --

[download]

As the regex never changes, there is no need to recompile it every time you call the sub, so I made it a constant. I would usually put the use constant.. inside the sub where it is used to keep it tidy, but I got beaten up because it implies that the constant is lexically scoped which it isn't. This doesn't fool me, but it is your choice. 

use constant RE_WORDS => qr[(?:[^\w\'\$!,.-]|(?:(?<=\D)[.,])|(?:[.,](?
+=\D|$)))+];

sub tokenize_msg_w_oneregex {
    my %words;
    @words{ split RE_WORDS, shift } = ();
    return keys %words;
}

[download]

To be fair, a large part of the savings is avoiding the map and initialising every hash element to 1 when you will never use the value. Doing the split inside a hash slice avoids this. You can easily feed this saving back into your other subs which would probably make your strings sub quickest again. But I thought I'd leave that AAEFTR:)

Replies are listed 'Best First'.
Re: Re: Re: Re: tokenize plain text messages
by revdiablo (Prior) on May 10, 2003 at 04:13 UTC

    Your suspicions are confirmed. The optimization has sped my strings-based tokenizer up by 39%. Interestingly, with the expanded test string I'm using, my original slightly edged out the optimized version of your single-regex. Here are the results:

               Rate    lists oneregex  str_org  str_opt
lists    2347/s       --     -38%     -45%     -60%
oneregex 3807/s      62%       --     -10%     -35%
str_org  4237/s      81%      11%       --     -28%
str_opt  5882/s     151%      55%      39%       --

    [download]

    This optimization is definitely effective. Thanks very much.

    Oh. And here's the expanded test string if you want to play with it:

    my $msg = q{This, is, an, example. Keep $2.50, 1,500, and 192.168.1.1.
+ 
I want to work this thing out a LITTEL!!!!L BITH!!!!! MORE@@@@@@ with 
some,.unhapp.yword,combinations.and , a little .. bit of,, 
confusing,text hopefully @#@#@#@%#$57)#$*(#&)(*$ it will @#@][] work.}
+;

    [download]
[reply]
[d/l]
[select]

      A couple of things I notice once I played with the expanded testcase. Your strings version is capturing single spaces somewhere. And both your strings and my regex are letting single $'s through as well.

      I have a m/(..)/g version which avoids these two problems and is quicker than my split attempt, but still not as quick as your strings in it's current form. 

      our $RE_WORDS = qr[
    (?: \$? \d+ (?:[.,] \d+ )*)+
    |
    [\w\'!-]+
]xo;
sub tokenize_msg_w_m {
    my %words;
    @words{ shift =~ m[($RE_WORDS)]og } = ();
    return keys %words;
}

      [download]

      Results

      D:\Perl\test>257026
Regex:
$ | want | 192.168.1.1 | $57 | confusing | yword |
 hopefully | combinations | LITTEL!!!!L | a | of |
 bit | is | This | to | will | text | this | 1,500 |
 Keep | out | BITH!!!!! | it | example | work | unhapp |
 $2.50 | little | MORE | I | some | thing | with | and | an

Strings:
$ |  | want | 192.168.1.1 | $57 | confusing | yword |
 hopefully | combinations | LITTEL!!!!L | a | of |
 bit | is | This | to | will | text | this | 1,500 |
 Keep | out | BITH!!!!! | it | example | work | unhapp |
 $2.50 | little | MORE | I | some | thing | with | and | an

Match:
want | 192.168.1.1 | $57 | confusing | yword |
 hopefully | combinations | LITTEL!!!!L | a | of |
 bit | is | This | to | will | text | this | 1,500 |
 Keep | out | BITH!!!!! | it | example | work | unhapp |
 $2.50 | little | MORE | I | some | thing | with | and | an

         Rate   regex   match strings
regex   526/s      --    -13%    -26%
match   607/s     15%      --    -15%
strings 710/s     35%     17%      --

      [download]

      The other question that crossed my mind was, what happens if the text contains "Do 33% of people watch T.V.?" ?

      Examine what is said, not who speaks.
      "Efficiency is intelligent laziness." -David Dunham
      "When I'm working on a problem, I never think about beauty. I think only how to solve the problem. But when I have finished, if the solution is not beautiful, I know it is wrong." -Richard Buckminster Fuller
[reply]
[d/l]
[select]

        The lonely $ is actually a legitimate token, and it's from near the end of the test string: ...&)(*$ it will...

        The empty tokens (they're not really spaces, but actually empty) are an artifact of the now-let's-delete-the-commas series of regex. It's due to a lone , or . surrounded by non-constituent chars, which initially gets considered as a token, but then deleted (leaving the temporary token separator). I suppose I could add another regex to s/%+/%/g, or just grep {!/^$/} keys %words..... or something.

        Here's the string shortly before being split on the temporary token boundaries; this is how I discovered where the anomalies were coming from.

        This%is%an%example%Keep%$2.50%1,500%and%192.168.1.1%I%want%to%work%thi
+s%thing%out%a%LITTEL!!!!L%BITH!!!!!%MORE%with%some%%unhapp%yword%comb
+inations%and%%a%little%%bit%of%%confusing%text%hopefully%$57%$%$%it%w
+ill%work

        [download]

        As for the TV example, that'd be tokenized just as it seems: Do|33|of|people|watch|T|V. This may seem slightly wrong, but I'm not trying to be perfect here. The digit-surrounded comma and period has already given me enough trouble. Gotta draw the line somewhere. :)

[reply]
[d/l]
[select]

In Section Seekers of Perl Wisdom