1 SIMS 290-2: Applied Natural Language Processing Marti Hearst October 18, 2004

1

SIMS 290-2: Applied Natural Language Processing

Marti HearstOctober 18, 2004

2

How might we analyze email?

Identify different partsReply blocks, signature blocks

Integrate email with workflow tasksBuild a social network

Who do you know, and what is their contact info?Reputation analysis

– Useful for anti-spam too

3

Today

Email analysisSpam filtering

4

Recognizing Email Structure

Three tasks:Does this message contain a signature block?If so, which lines are in it?Which lines are reply lines?Three-way classification for each line

RepresentationA sequence of linesEach line has features associated with itWindows of lines important for line classification

Victor R. Carvalho & William W. Cohen, Learning to Extract Signature and Reply Lines from Email, in CEAS 2004.

5Victor R. Carvalho & William W. Cohen, Learning to Extract Signature and Reply Lines from Email, in CEAS 2004.






11

The Cost of SpamMost of the cost of spam is paid for by the recipients:

Typical spam batch is 1,000,000 spams

Spammer averages ~$250 commission per batch

Cost to recipients to delete the load of spam @ 2 seconds/spam, $5.15/hour:

$2,861

12

The Cost of Spam

Theft efficiency ratio of spammer:

profit to thief ------------------------ = ~10 %

cost to victims

10% theft efficiency ratio is typical in many other lines of criminal activity such as fencing stolen goods (jewelery, hubcaps, car stereos).

13

How to Recognize Spam?

What features and algorithms should we use?

14Adapted froms slide by Rohan Malkhare

Anti-spam ApproachesLegislationTechnology

White listing of Email addressesBlack Listing of Email addresses/domainsChallenge Response mechanismsContent Filtering

– Learning Techniques– “Bayesian filtering” for spam has got a lot of press, e.g.

“How to spot and stop spam”, BBC News, 26/5/2003http://news.bbc.co.uk/2/hi/technology/3014029.stm

“Sorting the ham from the spam”, Sydney Morning Herald, 24/6/2003http://www.smh.com.au/articles/2003/06/23/1056220528960.html

– The “Bayesian filtering” they are talking about is actually Naïve Bayes Classification


Research in Spam Classification

Spam filtering is really a classification problemEach email needs to be classified as either spam or not spam (“ham”)

W. Cohen (1996): RIPPER, Rule Learning SystemRules in a human-comprehensible format

Pantel & Lin (1998): Naïve-Bayes with words as features

Sahami, Dumais, Heckerman, Horvitz (1998): Naïve-Bayes with a mutual information measure to select features with strongest resolving powerWords and domain-specific attributes of spam used as features


Research in Spam ClassificationPaul Graham (2002): A Plan for spam

Very popular algorithm credited with starting the craze for Bayesian FiltersUses naïve bayes with words as features

Bill Yerazunis (2002): CRM114 sparse binary polynomial hashing algorithm

Very accurate (over 99.7% accuracy)Distinctive because of it’s powerful feature extraction techniqueUses Bayesian chain rule for combining weightsAvailable via sourceforge

Others have used SVMs, etc.

New work: First email and anti-spam conference just heldhttp://www.ceas.cc/papers-2004/

17Adapted froms slide by William Yerazunis

Yerazunis’ CRM114 Algorithm

Other naïve-bayes approaches focused on single-word features

CRM114 creates a huge number of n-grams and represents them efficiently

The goal is to create a LOT of features, many of which will be invariant over a large body of spam (or nonspam).

(The name is a reference to a program in Dr. StrangeLove)

Sparse Binary Polynomial Hashing and the CRM114 Discriminator, William S. Yerazunis, http://crm114.sourceforge.net/CRM114_paper.html


CRM114

1. Slide a window N words long over the incoming text

2. For each window position, generate a set of order-preserving sub-phrases containing combinations of the windowed words

3. Calculate 32-bit hashes of these order-preserved sub-phrases (for efficiency reasons)


Step 1: slide a window N words long over the incoming text. ex:

You can Click here to buy viagra online NOW!!!

Yields:





... and so on... (on to step 2)

CRM114 Feature Extraction Example


SBPH Example

Click Click hereClick toClick here toClick buyClick here buyClick to buyClick here to buy Click viagraClick here viagraClick to viagraClick here to viagra Click buy viagraClick here buy viagraClick to buy viagraClick here to buy viagra

...yields all these feature sub-phrases

Note the binary counting pattern; this is the ‘binary’ in ‘sparse binary polynomial hashing’

Sliding Window Text : ‘Click here to buy viagra’

Step 2: generate order-preserving sub-phrases from the words in each of the sliding windows


SBPH Example

Click Click hereClick toClick here toClick buyClick here buyClick to buy Click here to buyClick viagraClick here viagraClick to viagraClick here to viagra Click buy viagraClick here buy viagraClick to buy viagraClick here to buy viagra

Step 3: make 32-bit hash value “features” from the sub-phrases

32-bit hash

E06BF8AA12FAD10F7B37C4F9113936CF1821F0E846B99AADB7EE69BF19A78B4D56626838AE1B0B615710DE7333094DBB

..... and so on


How to use the terms

For each phrase you can buildKeep track of how many times you see that phrase in both the spam and nonspam categories.

When you need to classify some text, Build up the phrases

– Each extra word adds 15 features

Count up how many times all of the phrases appear in each of the two different categories. The category with the most phrase matches wins.

– But really it uses the Bayesian chain rule


Learning and Classifying

Learning: each feature is bucketed into one of two bucket files ( spam or nonspam)

Classifying: the comparable bucket counts of the two files generate rough estimates of each feature's ‘spamminess’

P(F|C) =0.5 + ( |Fc| - |F~c| ) / ( 2 * MaxF )


The Bayesian Chain Rule (BCR)

P ( F|C ) P ( C )P (C|F ) = ------------------------------------------ P( F|C ) P( C ) + P ( F|~C) P(~C)

Start with P(C ) = P(~C) = .5For a new msg, compute this for both P(spam) and P(not-spam)Which ever has the higher score wins.

The denominator renormalizes to take into account if most of the email is mainly one class or the other


The feature set created by the SBPH feature hash gives better performance than single-word Bayesian systems.

Phrases in colloquial English are much more standardized than words alone - this makes filter evasion much harder

A bigger corpus of example text is better

With 400Kbytes selected spams, 300Kbytes selected nonspams trained in, no blacklists, whitelists, or other shenanigans

Evaluation


>99.915 % The actual performance of CRM114 Mailfilter from Nov 1 to Dec 1, 2002.

5849 messages, (1935 spam, 3914 nonspam)

4 false accepts, ZERO false rejects, (and 2 messages I couldn't make head nor tail of).

All messages were incoming mail 'fresh from the wild'. No canned spam.

For comparison, a human* is only about 99.84% accurate in classifying spam v. nonspam in a “rapid classification” environment.

Results


Filtering speed: classification: about 20Kbytes per second, learning time: about 10Kbytes per second (on a Transmeta 666 MHz laptop)

Memory required: about 5 megabytes

404K spam features, 322K nonspam features

Results Stats


The bad news: SPAM MUTATES

Even a perfectly trained Bayesian filter will slowly deteriorate.

New spams appear, with new topics, as well as old topics with creative twists to evade antispam filters.

Downsides?

29

Revenge of the Spammers

How do the spammers game these algorithms?Break the tokenizer

– Split up words, use html tags, etc

Throw in randomly ordered words– Throw off the n-gram based statistics

Use few words– Harder for the classifier to work

On Attacking Statistical Spam Filters. Gregory L. Wittel and S. Felix Wu, CEAS ’04.

30

Next Time

In-class work: creating categories for the Enron email corpus

Documents

1 SIMS 290-2: Applied Natural Language Processing Marti Hearst October 18, 2004