merlyn++. Neat! It's even a lot quicker than I thought it would be.

$text = do{ 
    local $/; 
    open W, '<dutch.words' or warn $!; 
    <W> 
};

print length $text;
2603063

use Benchmark qw[ cmpthese ];

cmpthese( 1, { 
    std=> q[ 
        @words1 = $text =~ m[\b(\w+)\b]g; 
    ],
    merlyn=> q[ 
        @words2 = $text =~ m[
            (?: (?<!\w)(?=\w) | (?<=\w)(?!\w) )
            (\w+)
            (?: (?<!\w)(?=\w) | (?<=\w)(?!\w) )
        ]xg 
    ] 
});
   (warning: too few iterations for a reliable count)
   (warning: too few iterations for a reliable count)

       s/iter merlyn    std
merlyn   25.9     --   -50%
std      13.0    98%     --

print scalar @words1, ' : ', scalar @words2;
227248  :  227248
[download]

...proving your assertion wrong.

It could be argued that my assertion that "that \b is neither a lookahead nor a lookbehind assertion, try creating one of either" is correct... in that.

• It isn't either, it's both!
• And your neat construct is actually two of each:) and I said "...one of either...".

but that would probably be a fruitless argument, and takes nothing away from your construction. Once again, it's neat :)