I've been comparing notes with Tom Yager, who notices lately that spammers' use of nonsense words, especially in Subject: headers, seems to be effective against the Bayesian filter in OS X's Mail.app. I checked, and SpamBayes is (so far) unaffected by this ploy. One of the cool things about SpamBayes is its ability to reveal how it analyzes messages. See below for its take on a message that has the Subject: line "Jon nezinyunyane inflechies" and a bunch of angle-bracketed garbage in the text.
Evidently SpamBayes ignores all the garbage tags, but since these serve as word delimiters, it winds up seeing a bunch of word fragments -- like 'innov' and 'ative' -- which it finds suspicious. And as the spam counts indicate, it has seen these fragments before, so over time their discriminatory power should only grow, not diminish.
It's also fascinating to look at the handling of the giveaway phrase "Multi-Trillion Dollar Market." "Multi-Trillion" does not appear on SpamBayes' list of interesting tokens, though "dollar" and "market" do. That SpamBayes makes no effort to correlate these adjacent words seems like an obvious limitation, and yet it is (so far) continuing to perform spectactularly well for me despite that.
Update: As I keep forgetting for some reason, and as Giorgio Valoti reminds me, Mail.app uses latent semantic analysis, not the Bayesian technique.
Spam Score: 1
word spamprob #ham #spam '*H*' 0 - - '*S*' 1 - - 'jon-' 0.0313807 38 1 'from:addr:jonathan' 0.06584 3 0 'noheader:mime-version' 0.267816 3682 1332 'there' 0.357648 1865 1027 'web' 0.359379 1678 931 'noheader:reply-to' 0.398404 8311 5444 'reply-to:none' 0.398404 8311 5444 'your' 0.607781 3493 5354 'now' 0.609287 1198 1848 'header:Date:1' 0.614892 5565 8789 'header:From:1' 0.616075 5536 8787 'live' 0.617098 227 362 'subject:Jon' 0.628519 123 206 "you've" 0.635875 294 508 'potential' 0.637791 171 298 'header:Received:6' 0.639839 738 1297 'url:com' 0.643368 3651 6515 'must' 0.651722 330 611 'year.' 0.654359 135 253 'proto:http' 0.657505 4086 7759 'area' 0.657698 183 348 'break' 0.668753 75 150 'skip:m 10' 0.671444 690 1395 'header:Return-Path:1' 0.698156 3807 8710 'join' 0.72811 152 403 'market.' 0.736942 76 211 'serious' 0.779225 64 224 'header:Message-Id:1' 0.782263 1220 4336 'sell' 0.789965 88 328 'life' 0.850081 110 618 'walls' 0.86821 15 99 'url:htm' 0.870883 144 962 'to:addr:jon' 0.871895 87 587 'url:index' 0.877064 152 1074 'dollar' 0.917399 19 211 'url:173' 0.934783 0 3 'unique,' 0.949503 5 97 'subject:\\xe9' 0.95032 1 23 'pac' 0.96723 3 94 'independence' 0.969427 3 101 'earn' 0.969434 11 352 '000' 0.971088 3 107 'url:61' 0.974407 1 46 'url:133' 0.97619 0 9 'margin' 0.977151 2 94 '$100,' 0.977616 2 96 'url:41' 0.981928 0 12 'ing' 0.982844 2 126 'infor' 0.987666 1 97 'ailability.' 0.994822 0 43 'rofit' 0.994822 0 43 'ative' 0.994938 0 44 'innov' 0.994938 0 44 'aking' 0.99505 0 45 'busin' 0.99505 0 45 'ited' 0.99505 0 45 'pture' 0.99505 0 45 'azing.' 0.995156 0 46 'lim' 0.995156 0 46 'ame' 0.99545 0 49 'che' 0.99545 0 49 'message-id:@vampiress.zzn.com' 0.995627 0 51 'from:addr:vampiress.zzn.com' 0.99579 0 53 'rica' 0.996014 0 56 'ess' 0.996937 0 73 'amearn' 0.997592 0 93 'amed' 0.997592 0 93 'fina' 0.997592 0 93 'kage.' 0.997592 0 93 'ncial' 0.997592 0 93 'ney!' 0.997592 0 93 'dre' 0.997667 0 96 'mation' 0.997738 0 99 'to:name:jon' 0.998132 0 120 'to:addr:songline.com' 0.998351 0 136
Message Stream:
Date: Thu, 10 Jul 2003 21:08:32 -0800 Subject: Jon nezinyunyane inflechies X-Sender: Cole Nickol <jonathan@vampiress.zzn.com> </pre> <pre> <ioIVkQkR><OlpyCsaP><MSvpk> <dHEnke><LmKTnRTowC><rrRsACDfT> <oxwCx><fRmtryOnGq><UUkBB> <XCKfvJ><cNPVNIa><qOaXN> <pJIdrYdnW><ipvub><TigRpgXSHL> <jOMnGnk><fAdDuDwhKg><vtWrFAnErq> <kSYDLiOeO><PMcHf><lmkJTSd> <JRFwXqJrU><aFFNaxP><PcdrQ> <interesses><nenahospodari><interesses> <p><font face="Trebuchet MS">Jon-</font></P> <p></p> <p> <nenahospodari><interesses></p> <p><font face="Trebuchet MS">Ca<NLsQcKSV>pture Your Dre<bSVhiD>amEarn Fina<interesses>ncial Independence</font></p> <p><font face="Trebuchet MS">You can now for the first time,<nenahospodari> <BJxQNXNYyQ>own a busin<Pqktbw>ess in your area with the most unique, <KvynRHwW>innov<FkEhUrwj>ative product in Ame<SatnmBdYiP>rica today. Work le<mQvuxWo>ss a week with the potential to earn $100,<interesses>000 a year. There is no sell<QgBARjeQ>ing and not ML<gtdyhs>M. Join a Multi-Trillion Dollar Market.</font></p> <p><font face="Trebuchet MS">The p<KChHi>rofit margin is am<SIGOVYIdkA>azing.<nenahospodari></font></p> <p><font face="Trebuchet MS"><nenahospodari>Break down the walls and live this life you've only dre<interesses>amed about.</font></p> <p><font face="Trebuchet MS">Lim<QkWkUgv>ited av<UPeOWb>ailability. for Your Fr<interesses>ee infor<RCbUkNPVg>mation pac<afkLWbCAa>kage.</font></p> <p><font face="Trebuchet MS"><oCgNBti><a href="http://61.173.41.133/3a22147895.com/index.htm">-web-site-</a></font> </p> <p><nenahospodari></p> <p><font face="Trebuchet MS">Y<rcMWMN>ou must che<bkvKm>ck this out if you are serious about m<TQYVM>aking mo<interesses>ney!</font></p> <nenahospodari> <p></p> <p></p> <p><font face="Trebuchet MS">O<interesses>pt o<yxXafJRXuQ>ut at web<interesses>site<interesses></font></p>
Former URL: http://weblog.infoworld.com/udell/2003/07/11.html#a741