Word salad (computer science)
From Wikipedia, the free encyclopedia
 |
This article needs additional citations for verification. Please help improve this article by adding reliable references. Unsourced material may be challenged and removed. (October 2007) |
 |
This article or section may contain original research or unverified claims. Please improve the article by adding references. See the talk page for details. (September 2007) |
Word salad is a mixture of seemingly meaningful words that together signify nothing;[1] the phrase draws its name from the common name for a symptom of schizophrenia, Word salad. When applied to a physical theory, "word salad" is a derogatory description that labels the theory as senseless or utterly devoid of meaning.[citation needed]
In the context of computer science and linguistics, explicitly constructed word salad is a tool for demonstrating the difference between random utterance and coherent expression of thought. Software such as the Dissociated press within emacs demonstrates the construction of interesting-but-meaningless word salad from large samples of coherent language, by constructing new, random documents that share some of the same word or letter clustering properties as the language sample. These word salads appear as natural language to the inattentive eye or ear, but are clearly meaningless when read or listened to with full attention. In the 21st century, spammers have begun using word salad construction as a way to elude e-mail filtering and attract web page indexing to spam.[2][3]
Contents |
[edit] Word salad with spam e-mail
In response to the growing problem of spam e-mail, filtering tools became available starting around 2002 which implemented a widely employed method known as the naive Bayes classifier. This method uses the probability of various words appearing in spam emails to automatically classify them as spam. For a short time, this worked fairly well to classify emails as probable spam. In response, spammers developed word salad to fool programs employing this method of classification.[4] By adding large amounts of random text somewhere in their message, spammers hope to confuse Bayesian classifiers into classifying the message as "ham e-mail" (non-spam e-mail). Typically, this text contains random words from a dictionary.[citation needed]
[edit] Word salad for web page spam
Gyöngyi and Garcia-Molina state this problem clearly:
"As more and more people rely on the wealth of information available online, increased exposure on the World Wide Web may yield significant financial gains for individuals or organizations. Most frequently, search engines are the entryways to the Web; that is why some people try to mislead search engines, so that their pages would rank high in search results, and thus, capture user attention."[2]
[edit] Sentence and paragraph salad
Paragraph salad will reduce the effectiveness of any of the algorithms mentioned above and will lead to higher scores with any Bayesian filters. The only algorithms that might thwart sentence and paragraph salad would be very high level and expensive natural language processing, some kind of artificial intelligence algorithm involving a search engine, or exhaustive listing of spam emails. All of these techniques would be exceptionally expensive, and would likely not be very successful at filtering spam despite their high cost.[citation needed]
In a related technique, actual text from some large corpus of legitimate English (the plays of Shakespeare, other etexts distributed by Project Gutenberg, random world wide web pages, Wikipedia, or the like) is added into the email. This approach attempts to get around algorithms that could be devised to detect the more primitive form of word salad.[5]
[edit] Letter salad
On an even smaller scale than word salad, spammers use misspellings of words to try to thwart Bayesian filters. Misspelling Viagra as Via6ra, \/|/\Gr/\, or any one of a number of other ways (see Leet), or even using characters from international character sets is an attempt to avoid the high efficiency with which a Bayesian filter would classify any email containing certain words as spam. A simple spell checker might significantly reduce the effectiveness of letter salad approaches, yet most present spam filters do not use one.[citation needed]
The lengths to which some spammers have gone with letter salad have often produced illegible, almost laughable messages. Reading such email has become akin to deciphering complex custom license plates.[citation needed]
[edit] Word salad filtering
Algorithms for detecting word salad are clearly possible and not particularly difficult to implement.[citation needed] They would be, for the most part, more computationally intensive than most rules used by spam filters today (2006). A statistical approach based on Zipf's law of word frequency has potential in detecting simple word salad, as do grammar checking and the use of natural language processing.[6] Statistical Markovian analysis, where short phrases are used to determine if they are likely to occur in normal English sentences, is another statistical approach that would be effective against completely random phrasing[6] but might be fooled by Dissociated press techniques.[citation needed]
[edit] Future
As spam filters get better at detecting simple word and letter salad, spammers will likely migrate towards sentence and paragraph salad techniques.[citation needed] In the process of obscuring their message from improving spam filters, they will also obscure their message from potential targets of their advertising, virus distribution, or phishing. At some point, the profitability of spam may be brought down to the point that its volume is substantially reduced.
[edit] Notes
[edit] References
Gyöngyi, Zoltán & Garcia-Molina, Hector (2005), “Web spam taxonomy”, Proceedings of the First International Workshop on Adversarial Information Retrieval on the Web (AIRWeb), 2005 in The 14th International World Wide Web Conference (WWW 2005) May 10, (Tue)-14 (Sat), 2005, Nippon Convention Center (Makuhari Messe), Chiba, Japan., New York, N.Y.: ACM Press, ISBN 1-59593-046-9
Lavergne, Thomas (2006). "Unnatural language detection". RJCRI'O6: Young Scientist' conference on Information Retrieval: 383-388, (French?). Retrieved on 2007-10-02.
