Information Extraction from the World Wide Web

1
Information Extractionfrom the World Wide Web

• Andrew McCallum
• University of Massachusetts Amherst
• William Cohen
• Carnegie Mellon University

2
Example The Problem
Martin Baker, a person
Genomics job
Employers job posting form
3
Example A Solution
4
Extracting Job Openings from the Web
5
Job Openings Category Food Services Keyword
Baker Location Continental U.S.
6
Data Mining the Extracted Job Information
7
What is Information Extraction
As a task
Filling slots in a database from sub-segments of
text.
October 14, 2002, 400 a.m. PT For years,
Microsoft Corporation CEO Bill Gates railed
against the economic philosophy of open-source
software with Orwellian fervor, denouncing its
communal licensing as a "cancer" that stifled
technological innovation. Today, Microsoft
claims to "love" the open-source concept, by
which software code is made public to encourage
improvement and development by outside
programmers. Gates himself says Microsoft will
gladly disclose its crown jewels--the coveted
code behind the Windows operating system--to
select customers. "We can be open source. We
love the concept of shared source," said Bill
Veghte, a Microsoft VP. "That's a super-important
shift for us in terms of code access. Richard
Stallman, founder of the Free Software
Foundation, countered saying
NAME TITLE ORGANIZATION
8
What is Information Extraction
As a task
Filling slots in a database from sub-segments of
text.
October 14, 2002, 400 a.m. PT For years,
Microsoft Corporation CEO Bill Gates railed
against the economic philosophy of open-source
software with Orwellian fervor, denouncing its
communal licensing as a "cancer" that stifled
technological innovation. Today, Microsoft
claims to "love" the open-source concept, by
which software code is made public to encourage
improvement and development by outside
programmers. Gates himself says Microsoft will
gladly disclose its crown jewels--the coveted
code behind the Windows operating system--to
select customers. "We can be open source. We
love the concept of shared source," said Bill
Veghte, a Microsoft VP. "That's a super-important
shift for us in terms of code access. Richard
Stallman, founder of the Free Software
Foundation, countered saying
IE
NAME TITLE ORGANIZATION Bill Gates
CEO Microsoft Bill Veghte VP
Microsoft Richard Stallman founder Free
Soft..
9
What is Information Extraction
As a familyof techniques
Information Extraction segmentation
classification clustering association
October 14, 2002, 400 a.m. PT For years,
Microsoft Corporation CEO Bill Gates railed
against the economic philosophy of open-source
software with Orwellian fervor, denouncing its
communal licensing as a "cancer" that stifled
technological innovation. Today, Microsoft
claims to "love" the open-source concept, by
which software code is made public to encourage
improvement and development by outside
programmers. Gates himself says Microsoft will
gladly disclose its crown jewels--the coveted
code behind the Windows operating system--to
select customers. "We can be open source. We
love the concept of shared source," said Bill
Veghte, a Microsoft VP. "That's a super-important
shift for us in terms of code access. Richard
Stallman, founder of the Free Software
Foundation, countered saying
Microsoft Corporation CEO Bill Gates Microsoft Gat
es Microsoft Bill Veghte Microsoft VP Richard
Stallman founder Free Software Foundation
10
What is Information Extraction
As a familyof techniques
Information Extraction segmentation
classification association clustering
October 14, 2002, 400 a.m. PT For years,
Microsoft Corporation CEO Bill Gates railed
against the economic philosophy of open-source
software with Orwellian fervor, denouncing its
communal licensing as a "cancer" that stifled
technological innovation. Today, Microsoft
claims to "love" the open-source concept, by
which software code is made public to encourage
improvement and development by outside
programmers. Gates himself says Microsoft will
gladly disclose its crown jewels--the coveted
code behind the Windows operating system--to
select customers. "We can be open source. We
love the concept of shared source," said Bill
Veghte, a Microsoft VP. "That's a super-important
shift for us in terms of code access. Richard
Stallman, founder of the Free Software
Foundation, countered saying
Microsoft Corporation CEO Bill Gates Microsoft Gat
es Microsoft Bill Veghte Microsoft VP Richard
Stallman founder Free Software Foundation
11
What is Information Extraction
As a familyof techniques
Information Extraction segmentation
classification association clustering
October 14, 2002, 400 a.m. PT For years,
Microsoft Corporation CEO Bill Gates railed
against the economic philosophy of open-source
software with Orwellian fervor, denouncing its
communal licensing as a "cancer" that stifled
technological innovation. Today, Microsoft
claims to "love" the open-source concept, by
which software code is made public to encourage
improvement and development by outside
programmers. Gates himself says Microsoft will
gladly disclose its crown jewels--the coveted
code behind the Windows operating system--to
select customers. "We can be open source. We
love the concept of shared source," said Bill
Veghte, a Microsoft VP. "That's a super-important
shift for us in terms of code access. Richard
Stallman, founder of the Free Software
Foundation, countered saying
Microsoft Corporation CEO Bill Gates Microsoft Gat
es Microsoft Bill Veghte Microsoft VP Richard
Stallman founder Free Software Foundation
12
What is Information Extraction
As a familyof techniques
Information Extraction segmentation
classification association clustering
October 14, 2002, 400 a.m. PT For years,
Microsoft Corporation CEO Bill Gates railed
against the economic philosophy of open-source
software with Orwellian fervor, denouncing its
communal licensing as a "cancer" that stifled
technological innovation. Today, Microsoft
claims to "love" the open-source concept, by
which software code is made public to encourage
improvement and development by outside
programmers. Gates himself says Microsoft will
gladly disclose its crown jewels--the coveted
code behind the Windows operating system--to
select customers. "We can be open source. We
love the concept of shared source," said Bill
Veghte, a Microsoft VP. "That's a super-important
shift for us in terms of code access. Richard
Stallman, founder of the Free Software
Foundation, countered saying
Microsoft Corporation CEO Bill Gates Microsoft Gat
es Microsoft Bill Veghte Microsoft VP Richard
Stallman founder Free Software Foundation




13
IE in Context
Create ontology
Spider
Filter by relevance
IE
Segment Classify Associate Cluster
Database
Load DB
Query, Search
Documentcollection
Train extraction models
Data mine
Label training data
14
Why IE from the Web?

• Science
• Grand old dream of AI Build large KB and reason
  with it. IE from the Web enables the creation
  of this KB.
• IE from the Web is a complex problem that
  inspires new advances in machine learning.
• Profit
• Many companies interested in leveraging data
  currently locked in unstructured text on the
  Web.
• Not yet a monopolistic winner in this space.
• Fun!
• Build tools that we researchers like to use
  ourselvesCora CiteSeer, MRQE.com, FAQFinder,
• See our work get used by the general public.

KB Knowledge Base
15
Tutorial Outline

• IE History
• Landscape of problems and solutions
• Parade of models for segmenting/classifying
• Sliding window
• Boundary finding
• Finite state machines
• Trees
• Overview of related problems and solutions
• Where to go from here

16
IE History

• Pre-Web
• Mostly news articles
• De Jongs FRUMP 1982
• Hand-built system to fill Schank-style scripts
  from news wire
• Message Understanding Conference (MUC) DARPA
  87-95, TIPSTER 92-96
• Most early work dominated by hand-built models
• E.g. SRIs FASTUS, hand-built FSMs.
• But by 1990s, some machine learning Lehnert,
  Cardie, Grishman and then HMMs Elkan Leek 97,
  BBN Bikel et al 98
• Web
• AAAI 94 Spring Symposium on Software Agents
• Much discussion of ML applied to Web. Maes,
  Mitchell, Etzioni.
• Tom Mitchells WebKB, 96
• Build KBs from the Web.
• Wrapper Induction
• Initially hand-build, then ML Soderland 96,
  Kushmeric 97,

17
What makes IE from the Web Different?
Less grammar, but more formatting linking
Newswire
Web
www.apple.com/retail
Apple to Open Its First Retail Store in New York
City MACWORLD EXPO, NEW YORK--July 17,
2002--Apple's first retail store in New York City
will open in Manhattan's SoHo district on
Thursday, July 18 at 800 a.m. EDT. The SoHo
store will be Apple's largest retail store to
date and is a stunning example of Apple's
commitment to offering customers the world's best
computer shopping experience. "Fourteen months
after opening our first retail store, our 31
stores are attracting over 100,000 visitors each
week," said Steve Jobs, Apple's CEO. "We hope our
SoHo store will surprise and delight both Mac and
PC users who want to see everything the Mac can
do to enhance their digital lifestyles."
www.apple.com/retail/soho
www.apple.com/retail/soho/theatre.html
The directory structure, link structure,
formatting layout of the Web is its own new
grammar.
18
Landscape of IE Tasks (1/4)Pattern Feature
Domain
Text paragraphs without formatting
Grammatical sentencesand some formatting links
Astro Teller is the CEO and co-founder of
BodyMedia. Astro holds a Ph.D. in Artificial
Intelligence from Carnegie Mellon University,
where he was inducted as a national Hertz fellow.
His M.S. in symbolic and heuristic computation
and B.S. in computer science are from Stanford
University. His work in science, literature and
business has appeared in international media from
the New York Times to CNN to NPR.
Non-grammatical snippets,rich formatting links
Tables
19
Landscape of IE Tasks (2/4)Pattern Scope
E.g. formatting layout patterns
Web site specific
Genre specific
Wide, non-specific
Amazon.com Book Pages
Resumes
Product Info
20
Landscape of IE Tasks (3/4)Pattern Complexity
E.g. word patterns
Regular set
Closed set
U.S. phone numbers
U.S. states
Phone [phone]
He was born in Alabama
The CALD main office can be reached at
412-268-1299
The big Wyoming sky
Ambiguous patterns,needing context andmany
sources of evidence
Complex pattern
U.S. postal addresses
Person names
University of Arkansas P.O. Box 140 Hope, AR
71802
was among the six houses sold by Hope Feldman
that year.
Pawel Opalinski, SoftwareEngineer at WhizBang
Labs.
Headquarters 1128 Main Street, 4th
Floor Cincinnati, Ohio 45210
21
Landscape of IE Tasks (4/4)Pattern Combinations
Jack Welch will retire as CEO of General Electric
tomorrow. The top role at the Connecticut
company will be filled by Jeffrey Immelt.
Single entity
Binary relationship
N-ary record
Person Jack Welch
Relation Person-Title Person Jack
Welch Title CEO
Relation Succession Company General
Electric Title CEO Out
Jack Welsh In Jeffrey Immelt
Person Jeffrey Immelt
Relation Company-Location Company General
Electric Location Connecticut
Location Connecticut
Named entity extraction
22
Evaluation of Single Entity Extraction
TRUTH
Michael Kearns and Sebastian Seung will start
Mondays tutorial, followed by Richard M. Karpe
and Martin Cooke.
PRED
Michael Kearns and Sebastian Seung will start
Mondays tutorial, followed by Richard M. Karpe
and Martin Cooke.
correctly predicted segments 2

Precision

predicted segments 6

correctly predicted segments 2

Recall

true segments 4
1
F1 Harmonic mean of Precision
Recall
((1/P) (1/R)) / 2
23
State of the Art Performance

• Named entity recognition
• Person, Location, Organization,
• F1 in high 80s or low- to mid-90s
• Binary relation extraction
• Contained-in (Location1, Location2)Member-of
  (Person1, Organization1)
• F1 in 60s or 70s
• Wrapper induction
• Extremely accurate performance obtainable
• Human effort (30min) required on each site

24
Landscape of IE Techniques (1/1)Models
Classify Pre-segmentedCandidates
Lexicons
Sliding Window
Abraham Lincoln was born in Kentucky.
Abraham Lincoln was born in Kentucky.
Abraham Lincoln was born in Kentucky.
member?
Classifier
Classifier
Alabama Alaska Wisconsin Wyoming
which class?
which class?
Try alternatewindow sizes
Context Free Grammars
Finite State Machines
Boundary Models
Abraham Lincoln was born in Kentucky.
Abraham Lincoln was born in Kentucky.
Abraham Lincoln was born in Kentucky.
BEGIN
Most likely state sequence?
NNP
V
P
NP
V
NNP
Most likely parse?
Classifier
PP
which class?
VP
NP
VP
BEGIN
END
BEGIN
END
S
and beyond
Any of these models can be used to capture words,
formatting or both.
25
LandscapeFocus of this Tutorial
Pattern complexity
closed set
regular
complex
ambiguous
Pattern feature domain
words
words formatting
formatting
Pattern scope
site-specific
genre-specific
general
Pattern combinations
entity
binary
n-ary
Models
lexicon
regex
window
boundary
FSM
CFG
26
Sliding Windows
27
Extraction by Sliding Window
GRAND CHALLENGES FOR MACHINE LEARNING
Jaime Carbonell School of Computer
Science Carnegie Mellon University
330 pm 7500 Wean
Hall Machine learning has evolved from obscurity
in the 1970s into a vibrant and popular
discipline in artificial intelligence during the
1980s and 1990s. As a result of its success and
growth, machine learning is evolving into a
collection of related disciplines inductive
concept acquisition, analytic learning in problem
solving (e.g. analogy, explanation-based
learning), learning theory (e.g. PAC learning),
genetic algorithms, connectionist learning,
hybrid systems, and so on.
CMU UseNet Seminar Announcement
28
Extraction by Sliding Window
GRAND CHALLENGES FOR MACHINE LEARNING
Jaime Carbonell School of Computer
Science Carnegie Mellon University
330 pm 7500 Wean
Hall Machine learning has evolved from obscurity
in the 1970s into a vibrant and popular
discipline in artificial intelligence during the
1980s and 1990s. As a result of its success and
growth, machine learning is evolving into a
collection of related disciplines inductive
concept acquisition, analytic learning in problem
solving (e.g. analogy, explanation-based
learning), learning theory (e.g. PAC learning),
genetic algorithms, connectionist learning,
hybrid systems, and so on.
CMU UseNet Seminar Announcement
29
Extraction by Sliding Window
GRAND CHALLENGES FOR MACHINE LEARNING
Jaime Carbonell School of Computer
Science Carnegie Mellon University
330 pm 7500 Wean
Hall Machine learning has evolved from obscurity
in the 1970s into a vibrant and popular
discipline in artificial intelligence during the
1980s and 1990s. As a result of its success and
growth, machine learning is evolving into a
collection of related disciplines inductive
concept acquisition, analytic learning in problem
solving (e.g. analogy, explanation-based
learning), learning theory (e.g. PAC learning),
genetic algorithms, connectionist learning,
hybrid systems, and so on.
CMU UseNet Seminar Announcement
30
Extraction by Sliding Window
GRAND CHALLENGES FOR MACHINE LEARNING
Jaime Carbonell School of Computer
Science Carnegie Mellon University
330 pm 7500 Wean
Hall Machine learning has evolved from obscurity
in the 1970s into a vibrant and popular
discipline in artificial intelligence during the
1980s and 1990s. As a result of its success and
growth, machine learning is evolving into a
collection of related disciplines inductive
concept acquisition, analytic learning in problem
solving (e.g. analogy, explanation-based
learning), learning theory (e.g. PAC learning),
genetic algorithms, connectionist learning,
hybrid systems, and so on.
CMU UseNet Seminar Announcement
31
A Naïve Bayes Sliding Window Model
Freitag 1997
00 pm Place Wean Hall Rm 5409
Speaker Sebastian Thrun


w t-m
w t-1
w t
w tn
w tn1
w tnm
prefix
contents
suffix
P(Wean Hall Rm 5409 LOCATION)
Prior probabilityof start position
Prior probabilityof length
Probabilityprefix words
Probabilitycontents words
Probabilitysuffix words
Try all start positions and reasonable lengths
Estimate these probabilities by (smoothed)
counts from labeled training data.
If P(Wean Hall Rm 5409 LOCATION) is above
some threshold, extract it.
Other examples of sliding window Baluja et al
2000 (decision tree over individual words
their context)
32
Naïve Bayes Sliding Window Results
Domain CMU UseNet Seminar Announcements
GRAND CHALLENGES FOR MACHINE LEARNING
Jaime Carbonell School of Computer
Science Carnegie Mellon University
330 pm 7500 Wean
Hall Machine learning has evolved from obscurity
in the 1970s into a vibrant and popular
discipline in artificial intelligence during the
1980s and 1990s. As a result of its success and
growth, machine learning is evolving into a
collection of related disciplines inductive
concept acquisition, analytic learning in problem
solving (e.g. analogy, explanation-based
learning), learning theory (e.g. PAC learning),
genetic algorithms, connectionist learning,
hybrid systems, and so on.
Field F1 Person Name 30 Location 61 Start
Time 98
33
SRV a realistic sliding-window-classifier IE
system
Frietag AAAI 98

• What windows to consider?
• all windows containing as many tokens as the
  shortest example, but no more tokens than the
  longest example
• How to represent a classifier? It might
• Restrict the length of window
• Restrict the vocabulary or formatting used
  before/after/inside window
• Restrict the relative order of tokens
• Etc

A token followed by a 3-char numeric token just
after the title
lttitlegtCourse Information for CS213lt/titlegt lth1gtCS
213 C Programminglt/h1gt
34
SRV a rule-learner for sliding-window
classification

• Top-down rule learning
• let RULES
• while (there are uncovered positive
  examples)
• // construct a rule R to add to RULES
• let R be a rule covering all examples
• while (R covers too many negative examples)
• let C argmaxC VALUE( R, R Æ C,
  uncoveredExamples)
• over some set of candidate conditions C
• let R R Æ C
• let RULES RULES R

35
SRV a rule-learner for sliding-window
classification

• Search metric SRV algorithm greedily adds
  conditions to maximize information gain of R
• VALUE(R,R,Data) IDatap ( p log p p log
  p)
• where p (p ) is fraction of data covered by R
  (R)
• To prevent overfitting
• rules are built on 2/3 of data, then their false
  positive rate is estimated with a Dirichlet on
  the 1/3 holdout set.
• Candidate conditions

36
Learning first-order rules

• A sample zero-th order rule set
• (tok1InTitle Æ tok1StartsPara Æ tok2triple)
• Ç (prevtok2EqCourse Æ prevtok1EqNumber) Ç
• First-order rules can be learned the same
  waywith additional search to find best
  condition
• phrase(X) Ã firstToken(X,A), startPara(A),
• nextToken(A,B), triple(B)
• phrase(X) Ã firstToken(X,A), prevToken(A,C),
  eq(C,number),
• prevToken(C,D), eq(D,course)
• Semantics
• p(X) Ã q(X),r(X,Y),s(Y) X 9 Y q(X) Æ
  r(X,Y) Æ s(Y)

37
SRV a rule-learner for sliding-window
classification

• Primitive predicates used by SRV
• token(X,W), allLowerCase(W), numerical(W),
• nextToken(W,U), previousToken(W,V)
• HTML-specific predicates
• inTitleTag(W), inH1Tag(W), inEmTag(W),
• emphasized(W) inEmTag(W) or inBTag(W) or
• tableNextCol(W,U) U is some token in the
  column after the column W is in
• tablePreviousCol(W,V), tableRowHeader(W,T),

38
SRV a rule-learner for sliding-window
classification

• Non-primitive conditions used by SRV
• every(X,W, f, c) 8 W2X f(W)c
• variables tagged must be used in earlier
  conditions
• underlined values will be replaced by constants,
  e.g., every(X, isCapitalized, true)
• some(X, W, ltf1,,fkgt, g, c) 9 W
  g(f1((fk(W)))c
• e.g., some(X, W, prevTok,prevTok,inTitle,false)
• set of paths ltf1,,fkgt considered grows over
  time.
• precedes(W,V), follows(W,V)
• tokenLength(X, relop, c)
• position(W,direction,relop, c)
• e.g., tokenLength(X,gt,4), position(W,fromEnd,lt,2)

39
Utility of non-primitive conditions in greedy
rule search

• Greedy search for first-order rules is hard
  because useful conditions can give no immediate
  benefit
• phrase(X) Ã token(X,A), prevToken(A,B),inTitle(B
  ),
• nextToken(A,C), tripleton(C)

40
Rapier an alternative approach
Califf Mooney, AAAI 99

• A bottom-up rule learner
• initialize RULES to be one rule per example
• repeat
• randomly pick N pairs of rules (Ri,Rj)
• let G1,GN be the pairwise generalizations
• let G argminG COST(G,RULES)
• let RULES RULES G R G R
• where COST(G,RULES) size of RULES- R G R
  and Gµ R means every example matching G matches
  R

41
lttitlegtCourse Information for CS213lt/titlegt lth1gtCS
213 C Programminglt/h1gt
courseNum(window1) Ã token(window1,CS),
doubleton(CS), prevToken(CS,CS213),
inTitle(CS213), nextTok(CS,213),
numeric(213), tripleton(213),
nextTok(213,C), tripleton(C), .
lttitlegtSyllabus and meeting times for Eng
214lt/titlegt lth1gtEng 214 Software Engineering for
Non-programmers lt/h1gt
courseNum(window2) Ã token(window2,Eng),
tripleton(Eng), prevToken(Eng,214),
inTitle(214), nextTok(Eng,214),
numeric(214), tripleton(214),
nextTok(214,Software),
courseNum(X) Ã token(X,A),
prevToken(A, B), inTitle(B),
nextTok(A,C)), numeric(C),
tripleton(C), nextTok(C,D),
42
Rapier an alternative approach

• Combines top-down and bottom-up learning
• Bottom-up to find common restrictions on content
• Top-down greedy addition of restrictions on
  context
• Use of part-of-speech and semantic features (from
  WORDNET).
• Special pattern-language based on sequences of
  tokens, each of which satisfies one of a set of
  given constraints
• lt lttok2ate,hit,POS2vbgt, lttok2thegt,
  ltPOS2nngtgt

43
Rapier results precision/recall
44
Rapier results vs. SRV
45
Rule-learning approaches to sliding-window
classification Summary

• SRV, Rapier, and WHISK Soderland KDD 97
• Representations for classifiers allow restriction
  of the relationships between tokens, etc
• Representations are carefully chosen subsets of
  even more powerful representations based on logic
  programming (ILP and Prolog)
• Use of these heavyweight representations is
  complicated, but seems to pay off in results
• Can simpler representations for classifiers work?

46
BWI Learning to detect boundaries
Freitag Kushmerick, AAAI 2000

• Another formulation learn three probabilistic
  classifiers
• START(i) Prob( position i starts a field)
• END(j) Prob( position j ends a field)
• LEN(k) Prob( an extracted field has length k)
• Then score a possible extraction (i,j) by
• START(i) END(j) LEN(j-i)
• LEN(k) is estimated from a histogram

47
BWI Learning to detect boundaries

• BWI uses boosting to find detectors for START
  and END
• Each weak detector has a BEFORE and AFTER pattern
  (on tokens before/after position i).
• Each pattern is a sequence of tokens and/or
  wildcards like anyAlphabeticToken, anyToken,
  anyUpperCaseLetter, anyNumber,
• Weak learner for patterns uses greedy search to
  repeatedly extend a pair of empty BEFORE,AFTER
  patterns

48
BWI Learning to detect boundaries
49
Problems with Sliding Windows and Boundary
Finders

• Decisions in neighboring parts of the input are
  made independently from each other.
• Naïve Bayes Sliding Window may predict a seminar
  end time before the seminar start time.
• It is possible for two overlapping windows to
  both be above threshold.
• In a Boundary-Finding system, left boundaries are
  laid down independently from right boundaries,
  and their pairing happens as a separate step.

50
Finite State Machines
51
Hidden Markov Models
HMMs are the standard sequence modeling tool in
genomics, speech, NLP,
Graphical model
Finite state model
S
S
S
transitions
t
-
1
t
t1
...
...
observations
...
Generates State sequence Observation
sequence
O
O
O
t
t
1
-
t
1
o1 o2 o3 o4 o5 o6 o7 o8
Parameters for all states Ss1,s2, Start
state probabilities P(st ) Transition
probabilities P(stst-1 ) Observation
(emission) probabilities P(otst ) Training
Maximize probability of training observations (w/
prior)
Usually a multinomial over atomic, fixed alphabet
52
IE with Hidden Markov Models
Given a sequence of observations
Yesterday Lawrence Saul spoke this example
sentence.
and a trained HMM
Find the most likely state sequence (Viterbi)
Yesterday Lawrence Saul spoke this example
sentence.
Any words said to be generated by the designated
person name state extract as a person name
Person name Lawrence Saul
53
HMM Example Nymble
Bikel, et al 1998, BBN IdentiFinder
Task Named Entity Extraction
Transitionprobabilities
Observationprobabilities
Person
end-of-sentence
P(ot st , st-1 )
P(st st-1, ot-1 )
start-of-sentence
Org
P(ot st , ot-1 )
or

(Five other name classes)
Back-off to
Back-off to
P(st st-1 )
P(ot st )
Other
P(st )
P(ot )
Train on 450k words of news wire text.
Case Language F1 . Mixed
English 93 Upper English 91 Mixed Spanish 90

Results
Other examples of shrinkage for HMMs in IE
Freitag and McCallum 99
54
Additional Work in HMMs for IE

• Early work
• Leek 1997, MS thesis with Elkans
• Back-off and Shrinkage
• Freitag McCallum 1999
• Active Learning
• Sheffer, 2001
• Learning Finite State Structure
• Freitag McCallum

Leave out this slide
55
Regrets from Atomic View of Tokens
Would like richer representation of text
multiple overlapping features, whole chunks of
text.

• line, sentence, or paragraph features
• length
• is centered in page
• percent of non-alphabetics
• white-space aligns with next line
• containing sentence has two verbs
• grammatically contains a question
• contains links to authoritative pages
• emissions that are uncountable
• features at multiple levels of granularity

• Example word features
• identity of word
• is in all caps
• ends in -ski
• is part of a noun phrase
• is in a list of city names
• is under node X in WordNet or Cyc
• is in bold font
• is in hyperlink anchor
• features of past future
• last person name was female
• next two words are and Associates

56
Problems with Richer Representationand a
Generative Model

• These arbitrary features are not independent
• Overlapping and long-distance dependences
• Multiple levels of granularity (words,
  characters)
• Multiple modalities (words, formatting, layout)
• Observations from past and future
• HMMs are generative models of the text
• Generative models do not easily handle these
  non-independent features. Two choices
• Model the dependencies. Each state would have
  its own Bayes Net. But we are already starved
  for training data!
• Ignore the dependencies. This causes
  over-counting of evidence (ala naïve Bayes).
  Big problem when combining evidence, as in
  Viterbi!

57
Conditional Sequence Models

• We would prefer a conditional modelP(so)
  instead of P(s,o)
• Can examine features, but not responsible for
  generating them.
• Dont have to explicitly model their
  dependencies.
• Dont waste modeling effort trying to generate
  what we are given at test time anyway.
• If successful, this answers the challenge of
  integrating the ability to handle many arbitrary
  features with the full power of finite state
  automata.

58
Locally Normalized Conditional Sequence Model
Maximum Entropy Markov Models McCallum, Freitag
Pereira, 2000 MaxEnt POS Tagger Ratnaparkhi,
1996 SNoW-based Markov Model Punyakanok Roth,
2000
Conditional
Generative (traditional HMM)
S
S
S
S
S
S
transitions
t
-
1
t
t1
transitions
t
-
1
t
t1
...
...
...
...
observations
observations
...
...
O
O
O
O
O
O
t
t
1
-
t
1
t
t
1
-
t
1
Standard belief propagation forward-backward
procedure. Viterbi and Baum-Welch follow
naturally.
59
Locally Normalized Conditional Sequence Model
Maximum Entropy Markov Models McCallum, Freitag
Pereira, 2000 MaxEnt POS Tagger Ratnaparkhi,
1996 SNoW-based Markov Model Punyakanok Roth,
2000
Or, more generally
Conditional
Generative (traditional HMM)
S
S
S
S
S
S
transitions
t
-
1
t
t1
transitions
t
-
1
t
t1
...
...
...
...
...
...
observations
...
entire observation sequence
O
O
O
O
t
t
t
1
-
t
1
Standard belief propagation forward-backward
procedure. Viterbi and Baum-Welch follow
naturally.
60
Exponential Form for Next State Function
st-1
Black-box classifier
weight
feature
Overall Recipe - Labeled data is assigned to
transitions. - Train each states exponential
model by maximum likelihood (iterative scaling
or conjugate gradient).
61
Feature Functions
o
Yesterday Lawrence Saul spoke this example
sentence.
o1 o2 o3
o4 o5 o6
o7
s1
s2
s3
s4
62
Experimental Data
38 files belonging to 7 UseNet FAQs
Example
ltheadgt X-NNTP-Poster NewsHound
v1.33 ltheadgt Archive-name acorn/faq/part2 ltheadgt
Frequency monthly ltheadgt ltquestiongt 2.6)
What configuration of serial cable should I
use? ltanswergt ltanswergt Here follows a
diagram of the necessary connection ltanswergt prog
rams to work properly. They are as far as I know
ltanswergt agreed upon by commercial comms
software developers fo ltanswergt ltanswergt
Pins 1, 4, and 8 must be connected together
inside ltanswergt is to avoid the well known
serial port chip bugs. The
Procedure For each FAQ, train on one file, test
on other average.
63
Features in Experiments

• begins-with-number
• begins-with-ordinal
• begins-with-punctuation
• begins-with-question-word
• begins-with-subject
• blank
• contains-alphanum
• contains-bracketed-number
• contains-http
• contains-non-space
• contains-number
• contains-pipe

• contains-question-mark
• contains-question-word
• ends-with-question-mark
• first-alpha-is-capitalized
• indented
• indented-1-to-4
• indented-5-to-10
• more-than-one-third-space
• only-punctuation
• prev-is-blank
• prev-begins-with-ordinal
• shorter-than-30

64
Models Tested

• ME-Stateless A single maximum entropy classifier
  applied to each line independently.
• TokenHMM A fully-connected HMM with four states,
  one for each of the line categories, each of
  which generates individual tokens (groups of
  alphanumeric characters and individual
  punctuation characters).
• FeatureHMM Identical to TokenHMM, only the lines
  in a document are first converted to sequences of
  features.
• MEMM The Maximum Entropy Markov Model described
  in this talk.

65
Results
66
From HMMs to MEMMs to CRFs
Conditional Random Fields (CRFs)
Lafferty, McCallum, Pereira 2001
St-1
St
St1
...
HMM
...
Ot
Ot1
Ot-1
MEMM
St-1
St
St1
...
Ot
Ot1
Ot-1
...
St-1
St
St1
...
CRF
Ot
Ot1
Ot-1
...
(A special case of MEMMs and CRFs.)
67
Conditional Random Fields (CRFs)
St
St1
St2
St3
St4
O Ot, Ot1, Ot2, Ot3, Ot4
Markov on s, conditional dependency on o.
Hammersley-Clifford-Besag theorem stipulates that
the CRFhas this forman exponential function of
the cliques in the graph.
Assuming that the dependency structure of the
states is tree-shaped (linear chain is a trivial
tree), inference can be done by dynamic
programming in time O(o S2)just like HMMs.
68
General CRFs vs. HMMs

• More general and expressive modeling technique
• Comparable computational efficiency
• Features may be arbitrary functions of any or all
  observations
• Parameters need not fully specify generation of
  observations require less training data
• Easy to incorporate domain knowledge
• State means only state of process, vsstate of
  process and observational history Im keeping

69
Efficient Inference
70
Training CRFs

• Methods
• iterative scaling (quite slow)
• conjugate gradient (much faster)
• conjugate gradient with preconditioning (super
  fast)
• limited-memory quasi-Newton methods (also super
  fast)
• Complexity comparable to standard Baum-Welch

Sha Pereira 2002 Malouf 2002
71
Voted Perceptron Sequence Models
Collins 2002
Like CRFs with stochastic gradient ascent and a
Viterbi approximation.
Analogous tothe gradientfor this onetraining
instance
Avoids calculating the partition function
(normalizer), Zo, but gradient ascent, not
2nd-order or conjugate gradient method.
72
MEMM CRF Related Work

• Maximum entropy for language tasks
• Language modeling Rosenfeld 94, Chen
  Rosenfeld 99
• Part-of-speech tagging Ratnaparkhi 98
• Segmentation Beeferman, Berger Lafferty 99
• Named entity recognition MENE Borthwick,
  Grishman,98
• HMMs for similar language tasks
• Part of speech tagging Kupiec 92
• Named entity recognition Bikel et al 99
• Other Information Extraction Leek 97, Freitag
  McCallum 99
• Serial Generative/Discriminative Approaches
• Speech recognition Schwartz Austin 93
• Reranking Parses Collins, 00
• Other conditional Markov models
• Non-probabilistic local decision models Brill
  95, Roth 98
• Gradient-descent on state path LeCun et al 98
• Markov Processes on Curves (MPCs) Saul Rahim
  99
• Voted Perceptron-trained FSMs Collins 02

73
Part-of-speech Tagging
Pereira 2001 personal comm.
45 tags, 1M words training data, Penn Treebank
DT NN NN , NN , VBZ
RB JJ IN PRP VBZ DT NNS , IN
RB JJ NNS TO PRP VBG NNS
WDT VBP RP NNS JJ , NNS
VBD .
The asbestos fiber , crocidolite, is unusually
resilient once it enters the lungs , with even
brief exposures to it causing symptoms that show
up decades later , researchers said .
Using spelling features
Error oov error error D err oov error D err
HMM 5.69 45.99
CRF 5.55 48.05 4.27 -24 23.76 -50
use words, plus overlapping features
capitalized, begins with , contains hyphen,
ends in -ing, -ogy, -ed, -s, -ly, -ion, -tion,
-ity, -ies.
74
Person name Extraction
McCallum 2001, unpublished
75
Person name Extraction
76
Features in Experiment

• Capitalized Xxxxx
• Mixed Caps XxXxxx
• All Caps XXXXX
• Initial Cap X.
• Contains Digit xxx5
• All lowercase xxxx
• Initial X
• Punctuation .,!(), etc
• Period .
• Comma ,
• Apostrophe
• Dash -
• Preceded by HTML tag

• Character n-gram classifier says string is a
  person name (80 accurate)
• In stopword list(the, of, their, etc)
• In honorific list(Mr, Mrs, Dr, Sen, etc)
• In person suffix list(Jr, Sr, PhD, etc)
• In name particle list (de, la, van, der, etc)
• In Census lastname listsegmented by P(name)
• In Census firstname listsegmented by P(name)
• In locations lists(states, cities, countries)
• In company name list(J. C. Penny)
• In list of company suffixes(Inc, Associates,
  Foundation)

Hand-built FSM person-name extractor says yes,
(prec/recall 30/95) Conjunctions of all
previous feature pairs, evaluated at the current
time step. Conjunctions of all previous feature
pairs, evaluated at current step and one step
ahead. All previous features, evaluated two steps
ahead. All previous features, evaluated one step
behind.
Total number of features 200k
77
Training and Testing

• Trained on 65469 words from 85 pages, 30
  different companies web sites.
• Training takes 4 hours on a 1 GHz Pentium.
• Training precision/recall is 96/96.
• Tested on different set of web pages with similar
  size characteristics.
• Testing precision is 0.92 - 0.95,
  recall is 0.89 - 0.91.

78
Inducing State-Transition Structure
Chidlovskii, 2000
K-reversiblegrammars
79
Limitations of HMM/CRF models

• HMM/CRF models have a linear structure
• Web documents have a hierarchical structure
• Are we suffering by not modeling this structure
  more explicitly?
• How can one learn a hierarchical extraction
  model?
• Coming up STALKER, a hierarchical
  wrapper-learner
• But first how do we train wrapper-learners?

80
Tree-based Models
81

• Extracting from one web site
• Use site-specific formatting information e.g.,
  the JobTitle is a bold-faced paragraph in column
  2
• For large well-structured sites, like parsing a
  formal language
• Extracting from many web sites
• Need general solutions to entity extraction,
  grouping into records, etc.
• Primarily use content information
• Must deal with a wide range of ways that users
  present data.
• Analogous to parsing natural language
• Problems are complementary
• Site-dependent learning can collect training data
  for a site-independent learner
• Site-dependent learning can boost accuracy of a
  site-independent learner on selected key sites

82
(No Transcript)
83
(No Transcript)
84
(No Transcript)
85
STALKER Hierarchical boundary finding
Muslea,Minton Knoblock 99

• Main idea
• To train a hierarchical extractor, pose a series
  of learning problems, one for each node in the
  hierarchy
• At each stage, extraction is simplified by
  knowing about the context.

86
(No Transcript)
87
(No Transcript)
88
(No Transcript)
89
(BEFORE(), AFTERnull)
90
(BEFORE(), AFTERnull)
91
(BEFORE(), AFTERnull)
92
Stalker hierarchical decomposition of two web
sites
93
Stalker summary and results

• Rule format
• landmark automata format for rules which
  extended BWIs format
• E.g. ltagtW. Cohenlt/agt CMU Web IE lt/ligt
• BWI BEFORE(lt, /, a,gt, ANY, )
• STALKER BEGIN SkipTo(lt, /, a, gt), SkipTo()
• Top-down rule learning algorithm
• Carefully chosen ordering between types of rule
  specializations
• Very fast learning e.g. 8 examples vs. 274
• A lesson we often control the IE training
  data!

94
Why low sample complexity is important in
wrapper learning
At training time, only four examples are
availablebut one would like to generalize to
future pages as well
95
Wrapster a hybrid approach to representing
wrappers
Cohen,JensenHurst WWW02

• Common representations for web pages include
• a rendered image
• a DOM tree (tree of HTML markup text)
• gives some of the power of hierarchical
  decomposition
• a sequence of tokens
• a bag of words, a sequence of characters, a node
  in a directed graph, . . .
• Questions
• How can we engineer a system to generalize
  quickly?
• How can we explore representational choices
  easily?

96
Wrapster architecture

• Bias is an ordered set of builders.
• Builders are simple micro-learners.
• A single master algorithm co-ordinates learning.
• Hybrid top-down/bottom-up rule learning
• Terminology
• Span substring of page, created by a predicate
• Predicate subset of spanspan, created by a
  builder
• Builder a micro-learner, created by hand

97
Wrapster predicates

• A predicate is a binary relation on spans
• p(s t) means that t is extracted from s.
• Membership in a predicate can be tested
• Given (s,t), is p(s,t) true?
• Predicates can be executed
• EXECUTE(s,t) t p(s,t)

98
Example Wrapster predicate
html

• http//wasBang.org/aboutus.html
• WasBang.com contact info
• Currently we have offices in two locations
• Pittsburgh, PA
• Provo, UT

head
body

p
p
WasBang.com .. info
ul
Currently..
li
li
a
a
Pittsburgh, PA
Provo, UT
99
Example Wrapster predicate

• Example
• p(s1,s2) iff s2 are the tokens below an li node
  inside a ul node inside s1.
• EXECUTE(p,s1) extracts
• Pittsburgh, PA
• Provo, UT

• http//wasBang.org/aboutus.html
• WasBang.com contact info
• Currently we have offices in two locations
• Pittsburgh, PA
• Provo, UT

100
Wrapster builders

• Builders are based on simple, restricted
  languages, for example
• Ltagpath p is defined by tag1,,tagk and
  ptag1,,tagk(s1,s2) is true iff s1 and s2
  correspond to DOM nodes and s2 is reached from s1
  by following a path ending in tag1,,tagk
• EXECUTE(pul,li,s1) Pittsburgh,PA, Provo,
  UT
• Lbracket p is defined by a pair of strings
  (l,r), and pl,r(s1,s2) is true iff s2 is preceded
  by l and followed by r.
• EXECUTE(pin,locations,s1) two

101
Wrapster builders

• For each language L there is a builder B which
  implements
• LGG( positive examples of p(s1,s2)) least
  general p2 L that covers all the positive
  examples (like pairwise generalization)
• For Lbracket, longest common prefix and suffix of
  the examples.
• REFINE(p, examples ) a set of ps that cover
  some but not all of the examples.
• For Ltagpath, extend the path with one additional
  tag that appears in the examples.
• Builders/languages can be combined
• E.g. to construct a builder for (L1 Æ L2) or (L1
  L2)

102
Wrapster builders - examples

• Compose tagpaths and brackets
• E.g., extract strings between ( and ) inside
  a list item inside an unordered list
• Compose tagpaths and language-based extractors
• E.g., extract city names inside the first
  paragraph
• Extract items based on position inside a rendered
  table, or properties of the rendered text
• E.g., extract items inside any column headed by
  text containing the words Job and Title
• E.g. extract items in boldfaced italics

103
Composing builders

• Composing builders for Ltagpath and Lbracket.
• LGG of the locations would be ptags pL,R where
• tags ul,li
• L (
• R )

• Jobs at WasBang.com
• Call (888)-555-1212 now to apply!
• Webmaster (New York). Perl, servlets essential.
• Librarian (Pittsburgh). MLS required.
• Ski Instructor (Vancouver). Snowboarding skills
  also useful.

104
Composing builders structural/global

• Jobs at WasBang.com
• Call Alberta Hill at 1-888-555-1212 now to apply!
• Webmaster (New York). Perl, servlets essential.
• Librarian (Pittsburgh). MLS required.
• Ski Instructor (Vancouver). Snowboarding skills
  also useful.

• Composing builders for Ltagpath and Lcity
• Lcity pcity where pcity(s1,s2) iff s2 is a
  city name inside of s2.
• LGG of the locations would be ptags pcity

105
Table-based builders
How to represent links to pages about
singers? Builders can be based on a geometric
view of a page.
106
Wrapster results
F1
examples
107
Wrapster results
Examples needed for 100 accuracy
108
Site-dependent vs. site-independent IE

• When is formatting information useful?
• On a single site, format is extremely consistent.
• Across many sites, format can vary widely.
• Can we improve a site-independent classifier
  using site-dependent format features? For
  instance
• Smooth predictions toward ones that are
  locally consistent with formatting.
• Learn a wrapper from noisy labels given by a
  site-independent IE system.
• First step obtaining features from the builders

109
Feature construction using builders

• - Let D be the set of all positive examples.
  Generate many small training sets Di from D, by
  sliding small windows over D.
• - Let P be the set of all predicates found by
  any builder from any subset Di.
• - For each predicate p, add a new feature fp that
  is true for exactly those x2 D that are extracted
  from their containing page by p.

110
List1
111
List2
112
List3
113
(No Transcript)
114
Learning Formatting Patterns On the
FlyScoped Learning
Bagnell, Blei, McCallum, 2002
Formatting is regular on each site, but there are
too many different sites to wrap. Can we get the
best of both worlds?
115
Scoped Learning Generative Model
a
q

• For each of the D documents
• Generate the multinomial formatting feature
  parameters f from p(fa)
• For each of the N words in the document
• Generate the nth category cn from p(cn).
• Generate the nth word (global feature) from
  p(wncn,q)
• Generate the nth formatting feature (local
  feature) from p(fncn,f)

f
c
w
f
N
D
116
Inference
Given a new web page, we would like to classify
each wordresulting in c c1, c2,, cn
This is not feasible to compute because of the
integral andsum in the denominator. We
experimented with twoapproximations - MAP
point estimate of f - Variational inference
117
MAP Point Estimate
If we approximate f with a point estimate, f,
then the integral disappears and c decouples. We
can then label each word with

A natural point estimate is the posterior mode a
maximum likelihood estimate for the local
parameters given the document in question
E-step
M-step
118
Global Extractor Precision 46, Recall 75
119
Scoped Learning Extractor Precision 58,
Recall 75 D Error -22
120
Broader View
Up to now we have been focused on segmentation
and classification
Create ontology
Spider
Filter by relevance
IE
Segment Classify Associate Cluster
Database
Load DB
Query, Search
Documentcollection
Train extraction models
Data mine
Label training data
121
Broader View
Now touch on some other issues
Create ontology
3
Spider
Filter by relevance
IE
Tokenize
Segment Classify Associate Cluster
1
2
Database
Load DB
Query, Search
Documentcollection
Train extraction models
4
Data mine
5
Label training data
1
122
(1) Association as Binary Classification
Sebastian Thrun conferred with Sue Becker, the
NIPS2002 General Chair.
Person
Person
Role
Person-Role (Sebastian Thrun, NIPS2002 General
Chair) ? NO
Person-Role ( Sue Becker, NIPS2002
General Chair) ? YES
Do this with SVMs and tree kernels over parse
trees.
Zelenko et al, 2002
123
(1) Association with Finite State Machines
Ray Craven, 2001
This enzyme, UBC6, localizes to the endoplasmic
reticulum, with the catalytic domain facing the
cytosol.
DET this N enzyme N ubc6 V localizes PREP to ART t
he ADJ endoplasmic N reticulum PREP with ART the A
DJ catalytic N domain V facing ART theN cytosol
Subcellular-localization (UBC6, endoplasmic
reticulum)
124
(1) Association using Parse Tree
Miller et al 2000
Simultaneously POS tag, parse, extract
associate!
Increase space of parse constitutes to
includeentity and relation tags Notation Descrip
tion . ch head constituent
category cm modifier constituent category Xp X of
parent node t POS tag w word Parameters e.g.
. P(chcp) P(vps) P
(cmcp,chp,cm-1,wp) P(per/nps,vp,null,said) P(tm
cm,th,wh) P(per/nnpper/np,vbd,said) P(wmcm,tm,t
h,wh) P(nanceper/np,per/nnp,vbd,said)
(This is also a great exampleof extraction using
a tree model.)
125
(1) Association with Graphical Models
Roth Yih 2002
Capture arbitrary-distance dependencies among
predictions.
126
(1) Association with Graphical Models
Roth Yih 2002
Also capture long-distance dependencies among
predictions.
Random variableover the class ofrelation
between entity 2 and 1, e.g. over lives-in,
is-boss-of,
person
Random variableover the class ofentity 1, e.g.
overperson, location,
lives-in
Local languagemodels contributeevidence to
relationclassification.
person?
Local languagemodels contributeevidence to
entityclassification.
Dependencies between classesof entities and
relations!
Inference with loopy belief propagation.
127
(1) Association with Graphical Models
Roth Yih 2002
Also capture long-distance dependencies among
predictions.
Random variableover the class ofrelation
between entity 2 and 1, e.g. over lives-in,
is-boss-of,
person
Random variableover the class ofentity 1, e.g.
overperson, location,
lives-in
Local languagemodels contributeevidence to
relationclassification.
location
Local languagemodels contributeevidence to
entityclassification.
Dependencies between classesof entities and
relations!
Inference with loopy belief propagation.
128
(1) Association of Entities from the Web
Web
EntitiesAssociatedinto Records
129
(1) Record association in Wrapster
5 label types sufficient for modeling 500 sites
Jensen Cohen, 2001
Toys.com Company InfoKitesBicycles
Name Box Kite Company Toys.comLocation
OregonOrder 1-800-FLY-KITECost
100Description Great for kidsColor
blueSize small
Company Info Location Oregon
Kites Box Kite 100 Stunt Kite 300
Name Stunt Kite Company Toys.comLocation
OregonOrder 1-800-FLY-KITECost
300Description Lots of funColor redSize big
Box Kite Great for kidsDetailed specs
Order Info Call1-800-FLY-KITE
Stunt Kite Lots of fun Detailed specs
Specs Color blueSize small
Specs Color redSize big
130
Broader View
Now touch on some other issues
Create ontology
3
Spider
Filter by relevance
IE
Tokenize
Segment Classify Associate Cluster
1
2
Database
Load DB
Query, Search
Documentcollection
Train extraction models
4
Data mine
5
Label training data
1
131
(2) Clustering for Reference Matching and
De-duplication
Borthwick, 2000
Learn Pr (duplicate, not-duplicate record1,
record2)with a Maximum Entropy classifier.
Do greedy agglomerative clustering using this
Probability as a distance metric.
132
(2) Clustering for Reference Matching and
De-duplication

• Efficiently clustering large data sets by
  pre-clustering with a cheap distance metric.
• McCallum, Nigam Ungar, 2000
• Learn a better distance metric.
• Cohen Richman, 2002
• Dont simply merge greedily capture dependencies
  among multiple merges.
• Pasula, Marthi, Milch, Russell, Shpitser,NIPS
  2002

133
Broader View
Now touch on some other issues
Create ontology
3
Spider
Filter by relevance
IE
Tokenize
Segment Classify Associate Cluster
1
2
Database
Load DB
Query, Search
Documentcollection
Train extraction models
4
Data mine
5
Label training data
1
134
(3) Automatically Inducing an Ontology
Riloff, 95
Two inputs
(1)
(2)
Heuristic interesting meta-patterns.
135
(3) Automatically Inducing an Ontology
Riloff, 95
Subject/Verb/Objectpatterns that occurmore
often in therelevant documentsthan the
irrelevantones.
136
Broader View
Now touch on some other issues
Create ontology
3
Spider
Filter by relevance
IE
Tokenize
Segment Classify Associate Cluster
1
2
Database
Load DB
Query, Search
Documentcollection
Train extraction models
4
Data mine
5
Label training data
1
137
(4) Training IE Models using Unlabeled Data
Collins Singer, 1999
says Mr. Cooper, a vice president of
NNP NNP