First, people *will* find ways around whatever filters you put in place. If you just want to stop the worst of the nasty words from showing up, you'll probably do OK at that, but your users will undoubtedly come up with new ways to say the same thing that won't be caught by your filter. What are you really trying to accomplish here? If it's a "letter of the law" situation, you've got half a chance. If you're trying to stop participants from communicating naughty ideas, you will fail.

Then, to make matters worse, the more words and permutations you try to filter out, the more false negatives you'll catch. Context is everything -- several years ago, AOL drew some bad publicity when breast cancer survivors were repeatedly dinged for using bad language in chat rooms and user profiles. That's one example. George Carlin sets forth several more in his "Seven Dirty Words" bit, like "You can prick your finger, but don't finger your prick" The more questionable words you try to block, the more legitimate conversation you will block unintentionally. And the more clever your users will get in their attempts to sidestep your bot.

Automating analysis of the English language is not something that can be done with a few perl regexps. Work on your bot, sure, but consider using it to alert a human who can read the questionable content in context and take action, rather than having the bot take action all by itself.

And once users realize that they are being filtered, some of them will find it a challange to get past the filter and so the filter may actually result in an increase in offensive chatting.

I recall a case where the source of the filter was identified and chat like "Bob, that is a big, stinky pile of Azhrarn and you know it" started showing up.

So one should probably make the filter very specific and rather simple so that it isn't much of an interesting challenge to get past and just catches the casual first use of a forbidden word as a reminder that such is not appreciated. This will catch the occasional "slip" by a well-meaning user while being less likely to inflame much of a response to the filter itself.

Though even that has a good chance of being more of a problem than it is worth.

        - tye (but my friends call me "Tye")

Tye has some very good points.

I run a couple of forums on which I've implemented a simple word list filter.

I find that I've backed off to just a few very obvious words. If one of these words show in text I replace with the Times Newspaper approved sustitute and also send myself a message.

Usually the substitution is enough to handle the immediate situation. If when the human (or semi-human) moi reviews the alerts a further action is required then further action is taken. But at this point the permutations of okay versus not okay are too complex for a simple bot to handle.

I've found that the simple substitution has sent a gentle message that certain language is not approved and most folks have either backed off or found a new and less offensive means of responding to each other.

In the few cases where the user was obviously way over the line I chose to implement a ban filter on that user. This was decided and implemented by the semi-human.

Claude

You could try "laundering" the text on its way in, rather than writing really complicated regexps. In other words, first render it in lowercase, remove all whitespace and punctuation, make substitions for l337 speak, then look for the nasties. It's fairly easy with tr:

$message = uc $message;          # uppercasify
$message =~ tr/ .;,/d;           # remove spaces and punctuation
$message =~ tr/13457/least/;     # un-1337
[download]

Do realize that the more munging you do in this fashion, the more prone your code will be to false positives. For example, the sample lines I wrote above would trigger on "I had amnesia" and "45 sweaters".

There are many issues with censorship, that'll try to avoid here, and I'd liek to extend the ideas already discussed in how to remove profanity

I think you'd get a lot of false matches on perfectly harmless language, so you need to gaurd against a lot of these without annoying people (IMHO, you'd annoy more people by punishing innocent than not stopping profanity)

The classic example I always use is my friend Dick from Scunthorpe. He has a pet Ass (as in donkey).

Now there are (at least) three false trigers in that paragraph (although some things like "c ex" might trigger more), and leads on to dialect; In UK English, an Ass is nothing but a donkey. You sit on your Arse, but getting drunk is being rat-arsed.

Perhaps on alternative is to match against bad, then match against that bad word's "safe" list (with words like Scunthorpe)

Ultimately there are going to be mistakes, and the only perfect way to do it would be human; although I'd be interested in developing an AI engine that could be taught...

--
RatArsed

Yeah, that was something like my second idea. If I washed the text, it would be a bit easier to match against a wordlist.
I don't really need to look for case though as //i gets rid of that. Also, I just thought of cases like multiple letters being substituted for one. Like "Ph" for "f." Would (ph|f) work on that for a word with the letter f in it? OR tr/(ph)/f/ ?
Hopefully not to many people have 45 sweaters. ;)

You just need a bigger wordlist. Since you're already going to need a list of bad words, why not put your leet speak words in there. Then you'll have fewer false matches on translated words like 45sweaters.

Anything that is suspect (isn't found in wordlist but contains numbers) could then be translated and run through a second matching.

It's all in your approach.

Check out the new Regexp::Common module that Damian Conway released, and IIRC japhy is now maintaining. It contains a built-in $RE{profanity} that you can match against without having to put the terminology in your script (and without having to try to define your own canon for what is "profane"). If you really wanted to, you could craft an extension (based on the way it's written) to the tree to implement the degree of comprehensiveness you desire. The paper describing it in detail is actually in the proceedings of TPC (which just ended Friday).

You may also want to watch how comprehensive you are trying to make this. Otherwise, you could find yourself filtering out things like i18n n d4m n l10n (for some legitimate use of "d4m").

Spud Zeppelin * spud@spudzeppelin.com

<soapbox=on mode="free speech">
[download]

I could likely sit here all day and come up with silly examples as to how this filtering scheme will filter more than you intended.

I could also show clever ways to still swear that most people would understand. I will not, however, profane the Monastery in that fashion.

What about historical references? How many of us, as we get ready to press the rest button and reinstall the server have shouted "damn the torpedoes, men!"? okay, I will admit I am likely to be in a minority on that one :)

My understanding is that most languages are far to context dependant for this kind of filter to work. I also freely admit this in no way helps answer your question, but I firmly believe there is no solution beyond compiling a list of Carlin's Seven Dirty Words and using a word boundary regex - like

  /\bdamn\b/
[download]

</soapbox>
[download]

Beware trying to automate any solution to the problems. Humans are simply too creative, and human language is simply too ambiguous.

I think your best bet is to monitor what your script considers to be the worst potential offenders, then have a pair of human eyeballs peruse the transcripts and take action when necessary.

I speak from experience: A dot-com I worked for tried to automatically filter user content on a site, and it monumentally backfired. There were so many embarrassing and problematic false positives that the filters were eventually removed.

(I should point out that the site wasn't even the kind one would think would generate such problems: It was a recreational sports site! As it turns out, alot of people like to include double entendres or clever wordplay when naming their sports teams ;-)

Hope this helps!

Interesting choice of terms considering what you want it to do. Some people could easily see "anal" as being offensive.

I suggest that you need to do some serious thinking about how much bang for the buck you need. As others have said, you're never going to get it worked out. Lord knows that on ddials years ago, it was easy to write "sh1t" and voila, you're past the filters.

How far are you willing to go here? How about spelling? Can your users discuss the song "If You See Kay" by Poster Children? (There was also a song with that title by April Wine)

How about the classic song from the Dr. Demento show "Polka Dot Undies"? How small are you willing to chase your diminishing returns?

xoxo,
Andy
--
<megaphone> Throw down the gun and tiara and come out of the float! </megaphone>

<point>Irregardless of what you try to filter, some people will get ar
+ound it, some will get filtered erroneously.  I would keep the list o
+f filtered words concise.
</point>
[download]

I probably will make it use word breaks, (but as far as I'm concerned Fucking, Austria would be a match ;) ) I'll have to see how sensitive word breaks are. As I would want punctuation to count, but wouldn't want it to trigger on any shroom discussions. ;)

Thanks for all the moral issues, but I will leave those dillemas up to the end users. I could also modify this once it is done to take ANY action, not just hostile.

I could make it trigger on key words like "Azhrarn" or "help" to alert me, and have it work out bad spellers and such. ;) This can be expanded quite a bit once it is completed, but it will be nice to have a base word-matching script that I can feed a list into. You guys just can't get past the basic premise of the script to other uses can you? :p

For me, at least, the issues I raised have less to do with morality than practicality (sorry to all you upvoters who thought I was making a grand anti-censorship speech :-) Parsing the English language is an extremely difficult task to automate, regardless of your motive or objective.

Unless you are using key words that are unlikely to be used in conversation unrelated to what you're trying to capture, you'll end up with bogus results unless you can also analyze and interpret context. Even humans don't do so well at that -- hence the use of "magic words" like "mayday" or "I am declaring an emergency" instead of "help" in situations where it really matters.

Trying to code this sort of understanding into a perl regex without resorting to predefined magic words would be quite a trick. Automating action based on anything a perl regex could identify simply won't do you much good in the long run.

This has been bugging me through the entire thread here. It could be that american chefs have managed to change the spelling we use to shittake, but the Japanese word is definatly shiitake, which is pronounced quite differently from shittake, now depending upon the way you want to romanize the word, shitake could be acceptable. I'm not personally familiar with all of the different romanization systems, but most of them romanize the japanese "‚µ‚¢‚½‚¯" as shiitake. As I said, I come at this as someone with a very basic understanding of the Japanese language, not as a chef. I'll stop rambiling now. -Ted

Required all words to be spelled properly. Then simply filter the words you want.

Heh, why didn't I think of that? :p Actually, even the words it does filter, will still be seen by the channel. It will just be a trigger for the bot to take action. There is no way to pass the text through the bot before it reaches the channel (unless you are using a botnet, but that is a different story).

Basically the bot will see someone say something it doesn't like and kick them off the channel as a warning. Although you can set a flag so it bans them instead. Or, as I said before, make it do just about anything.

The server automatically rejected any message which included the apparently profane phrase "Mass Spectrometry". Unfortunately, that's a pretty common phrase in analytical chemistry, and the site's maintainers never fixed the problem, so nobody ever used it. A sad waste of some poor programmer's time...

buckaduck