The information revolution has brought a new literary form: flames, or abusive electronic mail. As a researcher in information retrieval and classification, I analyzed 1,222 messages sent to the operators of controversial political sites on the World Wide Web in order to develop software to identify, and optionally filter out, messages that recipients would find insulting. Simple mechanical tricks worked surprisingly well, like looking for generalizations, imperative statements, and inflammatory political terms. Following are some of the more effective -- and amusing -- rules I discovered.
Flamers often insult not just the individual target of a message but the entire group that he or she represents to the flamer. I found that generalizations were much more likely to be insulting than flattering. Every message that contained “your ilk” was insulting, as in: “Your ilk is primarily responsible for most of the ills of this country.” (All quoted sentences are genuine.) Another insulting construct was “as/like yourselves": “Newt is a god untaintable by such pusilanimous [sic] vultures as yourselves.”
A related way of insulting someone was with “you” modified by a noun phrase (a noun apposition), like “you bozos,” “you flamers,” “you people,” and “you wacko liberals.” The only exceptions were “you guys” and “you folks,” and even those were three times as likely to occur in flames as in friendly messages.
“Scare quotes” and terms like “so-called” are used by writers to express skepticism about a subject. Looking for those quotes and terms helped find flames, as in the following message, received by Fairness & Accuracy in Reporting, a media-watch group that has criticized the accuracy of Rush Limbaugh’s statements: “The quotes you allegedly ‘debunk’ are taken out of context, your so-called ‘facts’ are merely opinions, and the tone of your articles are [sic] antagonistic.”
Short, imperative statements, in which the implicit subject is “you,” also tend to be insulting, such as: “Have your fun,” “Get lost,” and “Drop dead.”
Two common imperative sentences are not insulting, though: “Have a nice day” and “Keep up the good work.” Of the 70 messages containing the latter expression, 96 percent of them were friendly, making it an excellent predictor of nonflames.
Another class of imperative statements that are not insulting are requests, like “Please check us out and add us to your links!” Seventy-nine percent of the imperative sentences containing “please” were friendly. Two exceptions were “Please go crawl into a hole” and “Please extract your cranium from your rectum.”
Determining whether a sentence is imperative is easy for a human being but hard for the computer, which made many mistakes. For example, consider “Just saw our link,” sent by someone delighted to have discovered a link to his Web page. The computer interpreted the statement as a request to wield a handsaw. As a result of the computer’s difficulty in recognizing imperative statements, the knowledge that such statements tend to be insulting did not produce a useful way to detect flames automatically.
Whether a message is a flame can depend not just on its contents but on its recipient. The late S.I. Hayakawa, a linguist as well as a U.S. senator, noted that certain terms, which he called “snarl words,” are used only by those who wish to express disapproval. Thus, a message using snarl words to refer to the recipient’s beliefs is almost certainly a flame. Messages sent to the liberal Web site NewtWatch that contained “Slick Willie,” “Ted Kennedy,” or “socialist” were easily marked as hostile -- “You quivering socialist bedwetters must be scared to death” (also a noun apposition). Similarly, messages sent to conservative groups that mentioned “fascism” or “Watergate” also tended to be flames.
A unique method of insult used by liberals was to refer to someone’s rights, as in: “I respect and support your right to this conglomeration of crap but I had a difficult time keeping my lunch down as I browsed your pages.”
Using these guidelines and others, the software was able to correctly classify messages 95 percent of the time. (Sixty percent of the flames and 98 percent of nonflames were correctly identified.)
In the offline world, some people enjoy baiting others with intentionally offensive behavior, then enjoying their reactions. Online, this has become a sport called trolling for flames. Some trolls lack subtlety, like the one posted to the cat lovers’ discussion group that asked for recipes for cooking cats.
But others are deliciously clever. For example, in response to a discussion of the mystical and mathematical importance of ratios in the Great Pyramid of Cheops, John Baez posted a note that the ancient Greeks, too, were very sophisticated: If you take the ratio of the circumference to the radius of the columns they built, he wrote, you get an excellent approximation of pi. Gullible readers, not realizing that Baez was a mathematics professor with a sense of humor, protested that this was true of any circle. “That’s what makes it so spooky,” Baez cryptically replied, adding to the confusion.
The most common subjects of trolls are celebrities like L. Ron Hubbard and Ayn Rand, who have many vocal online supporters as well as detractors. Consider this message: “Ann [sic] Rand invented Game Theory during her love affair with Alan Turing -- the love affair, and the duel that ended his life, were dramatized by Harold Pinter in his play Rosencrantz and Turing Are Dead.”
While making an excellent piece of satire, this troll is too exaggerated to really fool anyone. Ayn Rand neither invented game theory nor knew the mathematician Alan Turing, who died of cyanide poisoning. Furthermore, most people know that the play’s real name is Guildencrantz and Rosenstern Are Dead, and that the author is Tom Stoppard. Experienced Internet users know that an obvious mistake is likely to be a plea for attention, and will resist the bait. An effective troll must be more subtle.
Ellen Spertus is an assistant professor of computer science at Mills College.
http://chronicle.com Section: Opinion & Arts Page: B6
