Smokey: Automatic Recognition of Hostile Messages
Contents
1. 0 Xclass ok Figure 1 Ordered rules generated by C4 5 Numbers in parentheses are rule numbers We also tried using linear regression which proved less successful The nominally independent variables were the 47 features plus 3 binary features indicating whether the message came from NewtWatch FAIR or The Right Side of the Web The dependent variable was 1 when the message was rated okay and 0 when rated a flame Messages for which there was disagreement were not used When a least squares analysis was performed on a subset of the 720 messages described earlier as the training data the resulting coefficients proved very accurate for the reserved portion but substantially less accurate for the test set correctly identifying 97 of the okay messages but only 39 of the flames We think the reason for the different performance is that the features had been tweaked to be consistent over the first set of 720 messages If an insulting adjective appeared in a message in this set and was not recognized as insulting it was manually added to the system The test set of 502 messages on the other hand was entirely out of sample analyzed only after the system had been frozen Besides its performance another disadvantage of linear regression is that it requires computing or at least bounding all of the features while the decision tree algorithm requires the computation of only some of the features Human classification okay maybe f2. occurring in real time communication in discussion groups such as Usenet newsgroups and mailing lists and in private messages such as email Publicly posted flames tend to be more clever and indirect than private email flames making them harder to reliably detect As a first step in this field Smokey addresses private messages specifically comments that are sent via feedback forms on World Wide Web pages In order to have enough flames to study I obtained messages from the webmasters of controversial pages specifically NewtWatch Dorsey and Schnackertz 1997 which criticizes Newt Gingrich The Right Side of the Web Donnels 1997 a conservative resource and Fairness and Accuracy in Reporting FAIR Ernst 1997 a media watch group best known for its criticisms of the veracity of Rush Limbaugh s claims The obvious method of identifying flames looking for obscene expressions does not work well Only 12 of the flames contained vulgarities and over a third of the vulgar messages were not flames Also some messages Copyright 1997 American Association for Artificial Intelligence www aaai org All rights reserved with profanity directed it at someone that both the sender and recipient dislike For example the statement Newt Gingrich is an a is a flame if sent to The Right Side of the Web but not if sent to NewtWatch as in fact it was All quoted examples and the typos in them are genuine The o
3. other way we overcame the generator s bias toward the common case was by interpreting its classifications of maybe as flame okay maybe flame okay flame original 574 88 58 9 9 1 removed 359 28 10 35 9 1 remaining 215 60 PE 4 5 1 Table 2 Messages of each type and okay flame ratio in original training set removed messages and remaining messages RESULTS C4 5 generated the rules shown in Figure 1 The results of the weeding and the rules on the test set are shown in Tables 3 and 4 Interpreting results of maybe as flames 98 of the okay messages were correctly classified as were 64 of the flames The machine classifications for messages that human volunteers disagreed on were considered to be don t cares reducing the size of the test set from 502 to 460 messages If Imperative short 13 gt 0 Condescension somewhat 21 lt 0 site specific insult 29 gt 0 Imperative short 13 lt 1 Insult recipient 25 gt 0 Insult other 28 gt 0 epithet 30 gt 0 Imperative short 13 lt 0 Profanity no villain 19 gt 0 Appositive guys 1 gt 0 site specific insult 29 gt 0 xclass flame If Appositive NP 3 lt 0 Insult villain 27 lt 0 site specific insult 29 lt 0 epithet 30 lt 0 exclamation points 47 lt 2 Appositive NP 3 lt 0 Imperative short 13 lt 0 site specific insult 29 lt 0 epithet 30 lt
4. 47 85 13 2 11 Imp containing love or like Love the artwork 9 100 0 0 12 Imp with comma or semicolon or more than 12 words Tf interested hit my page 63 62 21 17 13 Imp statement not meeting any of the above rules Get used to it 159 69 17 14 14 Sentence with a word beginning with you and with ilk Your ilk is primarily responsible for most of the ills of this 3 0 67 33 country 15 Sentence with your so called or your so called Read through your so called evidence 4 25 25 50 16 Sentence with as or like followed by yourself or Newt is a god untaintable by such pusilanimous vultures as 3 0 33 67 yourselves yourselves 17 Quoted phrase preceded by you or followed by of yours Why don t you fact check the news media 2 50 50 0 18 Sentence with obscene word and site specific villain or Newt Gingrich is an a is 4 100 0 0 19 Sentence with obscene word not meeting above rule What the f is your problem 17 29 12 59 20 Containing you miffed we you musn t mustn t If you are miffed about this election just wait till the next 3 0 0 100 chicken or what s the matter 21 Containing your right you have a right or bash While I respect your right to express opinions I feel that most 28 32 25 43 of what is posted here is 22 Conta
5. Smokey Automatic Recognition of Hostile Messages Ellen Spertus Microsoft Research MIT AI Lab and University of Washington Dept of Computer Science and Engineering Box 352350 Seattle WA 98195 ABSTRACT Abusive messages flames can be both a source of frustration and a waste of time for Internet users This paper describes some approaches to flame recognition including a prototype system Smokey Smokey builds a 47 element feature vector based on the syntax and semantics of each sentence combining the vectors for the sentences within each message A training set of 720 messages was used by Quinlan s C4 5 decision tree generator to determine feature based rules that were able to correctly categorize 64 of the flames and 98 of the non flames in a separate test set of 460 messages Additional techniques for greater accuracy and user customization are also discussed Introduction Flames are one of the current hazards of on line communication While some people enjoy exchanging flames most users consider these abusive and insulting messages to be a nuisance or even upsetting I describe Smokey a prototype system to automatically recognize email flames Smokey combines _ natural language processing and sociolinguistic observations to identify messages that not only contain insulting words but use them in an insulting manner Additional methods not implemented in Smokey are also outlined There are many different types of flames
6. a message as being a flame if it contained insulting or abusive language unless it was directed to someone the sender and recipient both disliked not merely if the sender expressed disagreement with the recipient Volunteers could classify a message with flame okay or maybe In 80 of the cases all four volunteers agreed In an additional 13 of the cases exactly three volunteers agreed I combined the ratings by classifying a message as a flame if at least three individuals considered it one 7 5 of the messages okay if at least 3 people judged it okay and nobody considered it a flame 80 and maybe otherwise 13 Message classification The decision tree generator C4 5 Quinlan 1993 was used with the MLC utilities Kohavi et al 1994 to generate a classifier Because decision tree generators perform badly when one classification is much more common than the others it was necessary to weed out messages that were obviously okay to lessen the imbalance This was done by observing that some features almost always indicated that a message was okay For example only of the 70 messages in the training set with keep up the good work was a flame By making the approximation that messages triggering any of 10 such rules italicized in Table 1 were okay the ratio of okay to flame in the remaining messages could be reduced from 10 1 to 4 5 1 as shown in Table 2 allowing effective decision tree generation The
7. age Proceedings of the Workshop Las Cruces New Mexico 1994 pp 96 103 Ruwet Nicolas Grammaire des Insultes et Autres Etudes Editions du Seuil Paris 1982 Sack Warren Representing and Recognizing Point of View Proceedings of the AAAI 95 Fall Symposium on AI Applications in Knowledge Navigation and Retrieval 1995 Trippett Frank Watching Out for Loaded Words in Seyler Dorothy U and Boltz Carol J eds Language Power pp 112 115 Random House New York 1986 Originally printed in Time May 24 1982 Rule Example n ok maybe flame 1 You followed by guys TIl be giving you guys a call soon 38 66 13 21 2 You followed by folks Well you folks have fun poking fun at Newt 8 63 25 13 3 You followed by any other noun phrase You quivering socialist bedwetters must be scared to death 30 33 13 53 to go to all this trouble bashing Newt 4 Imperative sentence Imp with have day Have a nice day 2 100 0 0 5 Imp with keep work or keep up Keep up the good work 112 97 2 1 6 Imp containing look Look forward to hearing from you 8 88 13 0 7 Imp containing take Take care 5 60 0 40 8 Imp containing let Let s not dilly dally 31 87 13 0 9 Imp containing thank Thank you 66 95 5 0 10 Imp containing please Please don t judge us all by our 6th district
8. ammar of insults in English Ruwet 1982 has written about the grammar of French insults Jochnowitz 1987 Quang 1992 and Allan and Burridge 1991 have written about idiomatic epithets in English There are numerous lists of and articles about dirty words see the bibliography in Jay 1992 or the publications of Maledicta Press Jay 1990 1992 has had students rate the offensiveness of various taboo words physiological responses to insults Dillard and Kinney 1994 have also been measured Hayakawa and Hayakawa 1990 and Trippett 1986 have written about the emotional content of terms particularly political ones Automatic categorization of texts has been a major area of information retrieval research Lewis 1992 Lewis and Hayes 1994 Sack has written a system to automatically determine the ideological bias of a text Sack 1995 Email classification through regular expressions is already in use such as through the mail program extensions Procmail Mailagent and Filter A different method of filtering unstructured text is through collaboration such as Tapestry Goldberg et al 1992 allowing users to rate individual pieces of bulk mail from mailing lists or news groups or individual senders other users can decide whether to read messages based on others appraisals The widely used LISTSERV list maintenance program L Soft 1995 includes a proprietary algorithm to detect spam inappropriately crossposted messages us
9. elphia John Benjamins 1992 Jochnowitz George Acceptable but not Grammatical Maledicta The International Journal of Verbal Aggression vol 9 1986 1987 pp 71 74 Kohavi Ron John George Long Richard Manley David and Pfleger Karl MLC A Machine Learning Library in C in Tools with Artificial Intelligence IEEE Computer Society Press 1994 pp 740 743 See http www sgi com Technology mlc L Soft International Inc List Owner s Manual for LISTSERV 1 8b http www lsoft com manuals ownerindex html December 12 1995 Lewis David Dolan Representation and Learning in Information Retrieval Doctoral Dissertation CS Department Univ Of Massachusetts Amherst 1992 Technical Report 91 93 Lewis David D And Hayes Philip J Guest Editorial ACM Transactions on Information Systems 12 3 July 1994 231 Maledicta Press P O Box 14123 Santa Rosa CA 95402 707 523 4761 Quang Phuc Dong English Subjects without Overt Grammatical Subject in Zwicky A M Salus P H Binnick R I amp Vanek A L eds Studies out in left field Defamatory essays presented to James D McCawley John Benjamins Philadelphia 1992 Quinlan J Ross C4 5 Programs for Machine Learning Morgan Kaufmann Publishers San Mateo CA 1993 Richardson Steve Bootstrapping Statistical Processing into a Rule based Natural Language Parser in The Balancing Act Combining Symbolic and Statistical Approaches to Langu
10. em Rule classes Noun Phrases used as Appositions I found that phrases with you modified by a noun phrase tend to be insulting The technical term is a noun apposition Examples are you bozos you flamers and you people Exceptions are you guys and sometimes you folks Table 1 shows that you guys appeared 38 times in the examined messages and that in the sentences it appeared 66 were in messages classified as okay 13 as maybe and 21 as flames You folks was less likely to be part of a flame 13 and the general case was more likely 53 Because the parser marks noun appositions in the grammatical tree of each sentence the Lisp rule to recognize them is trivial Unfortunately the parser sometimes misidentifies noun appositions in part because of typographical errors in the input such as T here are many other fine members of congress deserving of you gentile sympathies also Here the sender presumably meant to write your instead of you Imperative Statements Another heuristic based on the syntax of a sentence is that imperative statements commands tend to be insulting Some examples are Have your fun forget about it Get used to it Get over it Get Lost get a life f you but see Gregersen 1977 and Quang 1992 The guard for the class checks whether the parser marked the statement as being impe
11. ements or vulgarities others required the epithet rule such as MAYBE YOU SHOULD GET A LIFE AND QUIT TRYING TO USE RUSH AS YOUR WAY TO STARDOM This rule proved to be one of the most useful 18 of the flames included epithets and none of the non flames included them Polite Rules The politeness rule class is entered if a sentence does not contain an obscene word A message is considered polite Brown and Levinson 1987 if it contains thank but not no thanks please or constructs with would such as Would you be willing to e mail me your logo Praise Rules The praise class is entered if a sentence does not contain an obscene word The simplest rules are that a sentence is considered praise if it contains such word stems as bless godspeed kudos or congra for congratulations and related misspellings Other rules require vocabulary information I predefine regular expressions web nouns page site etc good adjectives great super etc and good verbs enjoy agree etc Each site also has a regular expression page name representing the name of the page and common synonyms such as NewtWatch and Newt Watch Rules check whether one of the positive terms occurs near the word you or a synonym for a page name such as This is my favorite political page If the word like appears the rule chec
12. g your and yourself are insulting specifically sentences with your ilk your so called and scare quotes such as This service of yours reminds me of when I was in college and kids wrote similar comments on the bathroom walls about Reagan Profanity Rules This class is entered when a sentence contains an obscene word The list did not include damn and hell since they are used so frequently A distinction is made depending on whether the sentence also contains the name of a site specific villain For example for NewtWatch villains are Newt Gingrich Rush Limbaugh and would fall in this category The names of web browsers such as lynx are considered honorary villains so this rule catches Lynx currently s ts out the first time you try to page down Condescension Rules Condescending statements recognized through regular expressions were divided into three classes very somewhat and slightly condescending and are described in Table 1 The only structural rule used for condescending statements is to mark a class of tag phrases two word phrases consisting of a contraction followed by another word and a question mark such as It really is a helpless feeling when your side is solidly in the minority isn t it The regular expression for such a tag phrase is a ZA Z t a zA Z Insults The rule class Insu
13. ick Willy GET THE SOCIALISTS OUT OF MY POCKET 49 29 29 43 or socialis to liberals 30 Insulting epithet such as get a life anywhere in sentence maybe you should get A life 33 21 39 39 31 Sentence with no obscene words SNOW that contains Thanks for this service 266 92 5 2 thanks 32 SNOW with please Please expose this hypocracy 128 86 10 4 33 SNOW with would you I would P d Would you be willing to email me your logo 157 90 8 1 34 SNOW with bless or godspeed Godspeed Good Luck and Stay True 5 100 0 0 35 SNOW with congra or kudos This is a very useful page congrads 12 100 0 0 36 SNOW with a positive adjective near a site synonym You have got a great web site here 157 94 5 1 BL SNOW with a positive verb near a site synonym we really enjoy the NEWIWATCH 36 92 8 0 38 SNOW with you near a positive adjective You have a very good thing going keep it up 34 88 3 9 39 SNOW with T near a good verb a good adjective at the I was delighted to find The Right Side 220 88 9 3 beginning of a sentence or at the end of a short sentence Great to see a conservative page on the net 40 SNOW with add near link or pointer or with shall I am adding links to your homepage from mine 13 92 8 0 will recommend near a si
14. ining you have got to be you ve got to be you You have got to be joking 1 0 100 0 23 Sentence with tag phrase Tt really is a helpless feeling when your side is solidly in the 2 50 50 0 minority isn t it Table 1 part 1 Column n indicates the number of times the construct occurred in the 1222 messages The okay maybe and flame columns indicate the probability that a sentence meeting the rule was in a message of that class for all messages the probabilities are 53 11 and 7 respectively Bold face is used to indicate probabilities higher than these baselines Italicized rules were useful predictors of okay messages Table 1 continues on the next page Rule Example n ok maybe flame 24 Contains a negative word near a term for the site You will regret that you had anything to do with this 9 22 33 44 crappy home page 25 Contains a negative word near you You Sick idiotic liberals 19 16 11 74 26 Contains a negative word near this used as a pronoun What kind of crap is this 3 33 33 33 27 Contains a negative word and a site specific villain All the criticism of Newt here is quite idiotic 40 70 13 18 28 Contains a negative word but does not meet any of the above Pardon my lack of tact but this is the most pathetic 128 52 27 20 rules thing I believe I have ever seen 29 Contains a site specific insulting phrase such as Sl
15. istics 1993 Dillard James Price and Kinney Terry A Experiential and Physiological Responses to Interpersonal Influence Human Communication Research vol 20 no 4 June 1994 502 528 Donnels Jeff The Right Side of the Web 1995 1997 http www clark net pub jeffd index html Dorsey Matt and Schnackertz Ted Jr NewtWatch 1995 1997 http www cais com newtwatch Ernst Michael Fairness and Accuracy in Reporting 1995 1997 http www fair org fair Goldberg D Nichols D Oki B and Terry D Using collaborative filtering to weave an Information Tapestry Communications of the ACM 35 12 December 1992 pp 61 70 Gregersen Edgar A A Note on English Sexual Cursing Maledicta The International Journal of Verbal Aggression vol 1 1977 pp 261 268 Hayakawa S I and Hayakawa Alan R Language in Thought and Action Fifth edition Harcourt Brace Jovanovich 1990 Herring Susan Posting in a Different Voice Gender and Ethics in Computer Mediated Communication in Ess Charles ed Philosophical Perspectives in Computer Mediated Communication Albany SUNY Press 1995 Infante Dominic A and Wigley Charles J HI Verbal Aggressiveness An Interpersonal Model and Measure Communications Monographs Volume 53 March 1986 Jay Timothy B What are Fighting Words Presented at Eastern Psychological Association March 1990 Philadelphia PA Jay T B Cursing in America Philad
16. ks that it is being used as a verb The sentence Like your pages qualifies but not cool progressive resources like Newt Watch where like is used as a conjunction A message may indirectly offer praise in a sentence with the word TI before a positive verb such as i just found your web pages and I love it or with a positive adjective at the end of a clause near the beginning of a sentence Very interesting Another way to offer indirect praise is to write that one will add or recommend a link or page Even mentioning the word link in a message means it is almost certainly friendly Miscellaneous The remaining classes of rules have a single rule each i e the guard was the rule and check for smiley faces phone numbers uniform resource locators web addresses offers laughter and exclamation points All are binary except for exclamation points for which a count is returned Method Human Message Ratings Each of the 1222 messages was rated by four speakers of American English employed by Microsoft who were not otherwise involved in this research Each message was rated by two men and two women because gender differences in online Herring 1995 and offline Jay 1992 flaming have been observed No individual rated messages from more than one site because it was important to remember the intended recipient s political orientation Volunteers were told to mark
17. lame okay 422 98 34 43 10 36 maybe 6 1 11 14 8 29 flame 4 1 34 13 10 36 Table 3 Confusion matrix for test set Human classification okay flame lokay 422 98 10 36 flame maybe 10 2 18 64 Table 4 Collapsed confusion matrix for test set Discussion Smokey s Limitations One flame could not be recognized because the typography was unusual GETOVERIT The following flame also managed not to trigger any rules 4S a jelly nosed poodle stomping candy brained cow clump For the champion of American mythology he sure knows how to knock down a common law tradition which protects the middle class ellipsis in original While an additional heuristic suggested by Phil Leone would recognize the cascaded adjectives in the first sentence as an insulting structure the lack of an explicit subject makes the first sentence hard to interpret Other flames pass by using friendly phrases sarcastically such as Keep up your efforts because I see them as truly benign and pointless Others are not sarcastic but are the exceptions to rules consider the following message as annotated by Smokey Praise delight I m glad to see that your incessant name calling and whining hasn t stopped Praise delight As long as it continues I m glad to say Il remian on the winning side of politics Because the chance that a message that triggers the P
18. lts is guarded by a check for bad words which consists of bad verbs stink suck etc bad adjs bad lousy etc and bad nouns loser idiot etc Rules check whether the bad word appears near a name for the page as in Your page is a JOKE or near the word you as in You Sick idiotic liberals A tule checks for a bad word near this used as a pronoun in place of a noun such as What kind of crap is this Sentences where this is used as an adjective to modify a noun are not counted such as Not only is this country in a bad state A separate rule checks for insults containing a site specific villain A separate class of insults is site specific phrases These include names Watergate is only mentioned in flames to The Right Side of the Web derogatory nicknames Slick Willy and terms that are primarily used when insulting a specific group e g calling liberals commies or conservatives fascists Hayakawa and Hayakawa 1990 Epithets The final class of insulting statements is epithets short insulting phrases Allan and Burridge 1991 Jochnowitz 1987 For example get followed within ten characters by life lost real clue with it or used to it Other epithets are two word phrases such as drop dead While some of these are caught by the check for imperative stat
19. nly change I made is replacing letters of obscene words with dashes Smokey looks not only for insulting words and the context in which they are used but also for syntactic constructs that tend to be insulting or condescending such as imperative statements Smokey avoids misclassifying friendly messages by looking for praise requests and polite speech System Architecture Smokey consists of 5 phases 1 Messages are converted into a common format with one sentence per line and delimiters between messages This is done in Emacs Lisp 2 The text is run through a parser developed by the Microsoft Research Natural Language Processing Group 3 The output of the parser is converted by sed and awk scripts into Lisp s expressions 4 The s expressions are processed through rules written in Emacs Lisp producing a feature vector for each message This phase can also generate output showing where each rule applied 5 The resulting feature vectors are evaluated with simple rules produced by the decision tree generator C4 5 Quinlan 1993 I removed duplicate excessively long and meaningless messages someone randomly pressing keys from the collections this could be automated Steps 1 and 3 of Smokey are trivial and will not be further discussed in this paper The parser used in step 2 is discussed elsewhere Richardson 1994 and as appropriate in this paper The s expressions used by the fourth step to represent each sentence incl
20. raise delight rule is a flame is only 4 5 Smokey makes a reasonable but wrong evaluation Limitations to Flame Recognition While some limitations on automatic flame identification are due to current natural language recognition technology others are inherent Fluent human readers are sometimes unable to tell whether a given message is friendly or sarcastic More practical problems are recognizing sarcasm and innuendo and making sense of complex sentences and mistakes in grammar punctuation and spelling which are all too common in email Here are some examples Sarcasm Thank you for recognizing the power of Newt Keep up the good work This was sent to NewtWatch Grammar etc mistakes What on earth a BIGGOT like you is doing walking onthe face of earth Innuendo Only cowards cheats thieves and liars hide behind pseudonyms The program cannot infer that the sender is referring to the recipient who uses a pseudonym Fortunately statements that are meant to be insulting tend to be near other insults allowing a message to be correctly labeled even when individual sentences cannot be Possibilities for future work include learning from dictionaries and thesauri Dolan et al 1993 user feedback or proximity to known insults morphological analysis spelling and grammar correction and analyzing logical parse trees of sentences Related Work Surprisingly little has been written on the gr
21. rative There are several varieties of imperative statements that are not insulting including idiomatic expressions such as keep up the good work and have a nice day suggestions and invitations and sentences that only appear to be imperative because the writer omitted T such as Love your work Long imperative statements or those with multiple clauses are less likely to be insulting Consider If you have a candidate pledge your loyalty to only one and don t make a mistake and lose yourself in congress All of these conditions are checked for by rules in the imperative class as illustrated in Table 1 A source of miscategorizations is ambiguous statements such as the following Cool page Just saw our link The parser identified these as imperative statements a reasonable but amusing interpretation Cool page was marked as imperative because it can be interpreted as a verb followed by a noun presumably meaning that the listener should take a page and cool it off Just saw our link was misinterpreted as a command to saw i e with a handsaw a link Semantic analysis wouldn t necessarily help because sawing a link i e of a chain makes sense Of course neither sentence was meant as imperative so the rule misfires contributing to its lower than expected flame prediction rate Second Person Rules Many sentences with a word beginning with you includin
22. rston Deb Coughlin Bill Dolan Karen Jensen and Lucy Vanderwende for their assistance with natural language processing and linguistics to David Heckerman Carl Kadie John Miller and Eric Anderson for advice on decision theory and statistics and to David Lewis for invaluable encouragement and help with information retrieval techniques and the writing of this paper This work would not have been possible without the support of Daniel Weise Dan Ling and Rick Rashid and the help of Kevin Shields I received outstanding library support from Pamela Bridgeport and Keith Barland and technical support from Robert Eberl and Ken Martin I am grateful to the individuals and organizations that contributed email Jeff Donnels Matt Dorsey and Ted Schnackertz Michael Ernst and Mike Silverman I had valuable discussions with Reinhold Aman Cathy Ball Gene Ball Oren Etzioni Chris Gual David Kurlander Phil Leone Nate Osgood and David Perlmutter Timothy Jay provided me with the most recent version of his excellent bibliography References Allan Keith and Burridge Kate Euphemism amp Dysphemism Oxford University Press 1991 Brown Penelope and Levinson Stephen Politeness some universals in language usage Cambridge University Press 1987 Dolan William B Vanderwende L and Richardson S Automatically Deriving Structured Knowledge Base from On line Dictionaries in Proceedings of the Pacific Association for Computational Lingu
23. te synonym T shall use your page as a guide 41 SNOW containing link and not meeting any of the above I invite you to establish links to my newt bites And 82 99 1 0 rules Children Go Hungry sticker page 42 Contains a smiley face such as or I see where you are coming from Liberal losers who 18 83 6 11 can t get over the loss in 94 43 Contains a telephone number I am told that 1 800 768 2221 connects directly to the 24 92 4 4 Congressional switchboard free of charge 44 Contains a uniform resource locator URL i e a web You can check it out at http www sfgate com 127 98 1 1 45 Contains T near help or give if I can help let me know 12 83 17 0 46 Contains laughter Ha ha ha ha ha ha ha ha Your page is a joke 6 50 17 33 47 Contains exclamation points neWt gInGriCh iS evil 460 81 10 9 Table 1 part 2 Column n indicates the number of times the construct occurred in the 1222 messages The okay maybe and flame columns indicate the probability that a sentence meeting the rule was in a message of that class for all messages the probabilities are 83 11 and 7 respectively Bold face is used to indicate probabilities higher than these baselines Italicized rules were useful predictors of okay messages
24. ually advertisements If the software determines that a user has sent spam the message and subsequent ones from the same user will be sent to the list owner for approval combining automatic and social filtering Implications One advantage of mailbox filters such as Smokey is that they do not infringe on freedom of speech People are both free to write what they wish to willing readers and to not read anything they don t want to Assuming individuals can train their own filters nobody will be able to control what anybody else can read While Smokey isn t perfect it could be used now however to prioritize mail Maintainers of controversial web sites who are overwhelmed by mail could use it to move suspected non flames up in priority They could avoid suspected flames when busy or when already in a bad mood without delaying much inoffensive email Similar techniques could be used for other email related tasks such as eliminating unsolicited advertisements or routing mail sent to a general company address to the right individual A new arms race is starting As these and similar rules get published flamers will learn how to get around them Still there is a net benefit since the obscene expressions that affect people most emotionally can be eliminated and most flamers will not be knowledgeable about the defense systems especially if they are tailored to individuals Acknowledgments I am grateful to Diana Peterson Simon Co
25. ude 1 an ordinary string 2 a list of words in the sentence and 3 a tree encoding the grammatical structure of each sentence Each rule in step 4 is a Lisp procedure that takes an s expression representing a sentence as an argument and returns 1 if the rule applies O otherwise The only exception was the rule that returns a count of exclamation points Mutually exclusive rules are grouped into classes for greater accuracy and efficiency Each class has a guard procedure which checks whether a rule in the class could apply If so the rules are attempted in order until one succeeds All regular expression matching was case Innovative Applications of Artificial Intelligence IAAI 97 insensitive Table 1 shows each of the rules the number of times it is met in the sample data and the probability that if a sentence satisfies the rule that it is part of a message classified as a flame maybe or okay The reader will want to refer to Table 1 while reading the next section which describes the rule classes A feature vector was created for each message by summing the vectors of each sentence Most of the rules behave the same regardless of where the data comes from There are a few variables that get set with site specific information For example the string Slick Willy is probably insulting in a message to NewtWatch but not in one to The Right Side of the Web Site specific variables will be discussed with the rules that use th
Download Pdf Manuals
Related Search