I used to work at a publishing services company editing the text of books. If I was looking at a page with an editor's mark, and I wanted to find that place in the file, I'd look for a nearby unusual word or phrase and search for that in the file.

So my suggestion is to focus on the parts of each text that make them most unique. Take a word histogram of the original page text and look for the least common words in the quote. From there, you can look nearby in the original page and see if other low frequency words are nearby.

I can't tell from your problem description whether you're more concerned with false positives or false negatives. I think the algorithm I'm describing could be fairly tricky, and probably hard to tune to a particular false detection rate. Still, depending on the nature of the texts you have, it might pick up things other matching methods miss.

Update: Added a little more substance to the demonstration.

String::Approx and Text::Levenshtein may not be as useful as you you might expect for this as they operate on the string as an array of chars.

It's easy to see how "cunning stunt" and it's rude spoonerism would be rated as closely synonymous by such algorithms.

Perhaps a better approach, once you've extracted the raw text from the html, would be to search for the individual words from the quote, in the extracted text, and look for a high proportion of matches in close proximity.

If you stored the matches in a hash, wordPositionInText => word, then look for runs of consecutive, or nearly consecutive positions in the result, then you will find likely candidates easily. For example, match the edited quote "I want to get a little more sophisticated and allow the user to submit a mildly edited quote." against the text from your post:

#! perl -slw
use strict;
use List::Util qw[ reduce ];
use Data::Dump qw[pp];

sub normalise {
    local $_ = lc shift;
    tr[!"£$%^&*()-_=+'@;:#~/?\|`][]d; #"
    s[\s+][ ]g;
    return $_;
}

my $text = normalise do{ local $/; <DATA> };
my( $n, %wordPosns ) = 0;
$wordPosns{ pos( $text )-1 } = ++$n while $text =~ m[\s+]g;

my $quote = normalise 'I want to get a little more sophisticated and a
+llow the user to submit a mildly edited quote.';
my @qwords = split ' ', $quote;

my %matches;
for my $word ( @qwords ) {
    while( $text =~ m[\b$word\b]g ) {
        $matches{ $wordPosns{ pos $text } } = $word ;
    }
}
#pp \%matches;

my @runs = [];
reduce {
    push @{ $runs[ -1 ] }, $a;
    push @runs, [] if defined $a and $a+1 < $b;
    $b;
} sort{ $a <=> $b } keys %matches;

@runs = grep @$_ > 3, @runs;
#pp \@runs;

print join ' ', map $matches{ $_ }, @$_ for @runs;

__DATA__
[download]

produces:

C:\test>665700
i want to get a little more sophisticated
i want the user to
to submit a mildly edited quote
[download]

As usual, you rule. I'll play around with that.

Ah, how I love to see #" at the end of a line of someone else's code too. :)

That said, your desire to allow the user "to submit a mildly edited quote" reaches deep into the subjective... and thus into a province where human cognition still performs better than automated "judgment." BrowserUK's rude spoonerism and kind might be dealt with by adding a stop-list... but we've seen that battle fought and lost. And what might we do to deal with a case where the submitter modifies the original text (containing a word or phrase which is in itself unobjectionable) so that its meter, rhyme or other characteristic --- by virtue of the modification -- suggests an objectionable spoonerism?

For me, though, even bigger concerns arise in the area of accuracy and context. As we've seen recently in (US presidential) politics and elsewhere, an accurately quoted but out of context fragment of an utterance can turn the intent of the original upon its head. So too can what automata might well consider mild editing: for example: a quote which adds a negative or something on the order of s/"We have nothing to fear but fear itself."/"We have to fear fear itself."/)

Thus, though YMMV may vary with your intended application, /me believes that in many cases the tool you're building may -- given the state-of-the-art, today -- still require the application of human eyeballs and judgement in many or most cases.

Update: added "and to OP." /me has no excuse for omitting that in the original.

How about parsing the quote into grammatical pieces? You could then compare the parse trees for similarity in addition to looking at what text has changed.

If you can't parse the sentence at all, that's a good hint that it is spam. Otherwise, having a tree handy would make it easy to provide hints and highlights for the human editor to look at, and save their time.