This document proposes a framework called ALFRED for learning web wrappers from crowdsourced inputs. ALFRED uses a Bayesian model to evaluate candidate extraction rules based on worker responses to membership queries. It employs techniques like active learning to select the most informative queries. Additionally, it dynamically expands the expressiveness of the extraction language if needed, inspired by structural risk minimization, in order to find accurate rules while minimizing queries. The framework is evaluated on real datasets and is shown to significantly reduce the number of queries needed compared to static approaches, while maintaining high precision and recall.

1. A Framework for Learning Web
Wrappers from the Crowd
Valter Crescenzi, Paolo Merialdo, Disheng Qiu
Dipartimento di Ingegneria
Università degli Studi Roma Tre
Via della Vasca Navale, 79, Rome
disheng@dia.uniroma3.it

2. Extracting data
2M pages from IMDB, and we want to extract ... titles, directors etc ....
1/15

3. Extracting data
2M pages from IMDB, and we want to extract ... titles, directors etc ....
DB#Wrapper!
1/15

4. Extracting data
2M pages from IMDB, and we want to extract ... titles, directors etc ....
Inference
algorithm!
DB#Wrapper!
1/15

5. Supervised
Supervised hard to scale
Inference
algorithm!
DB#Wrapper!
1/15

6. Unsupervised
Unsupervised easier to scale but not accurate
Inference
algorithm!
DB#Wrapper!
1/15

7. Automatic Annotator
Automatic annotators can not be applied in all cases
Inference
algorithm!
DB#Wrapper!
+"
1/15
• Sample values
• Ontology
• Lexical patterns

8. Crowdsourcing
An opportunity to scale supervised approaches
Inference
algorithm!
DB#Wrapper!
1/15

9. Scaling Wrapper Inference
Scaling the number of workers with Crowdsourcing platforms opens new
challenges:
Issues: Contributions:
2/15

10. Scaling Wrapper Inference
Scaling the number of workers with Crowdsourcing platforms opens new
challenges:
Issues: Contributions:
Non-expert
workers
• Simple interactions to reduce the
worker error rate
• Membership Query (yes/no answer)
2/15

11. Scaling Wrapper Inference
Scaling the number of workers with Crowdsourcing platforms opens new
challenges:
Issues: Contributions:
Non-expert
workers
• Simple interactions to reduce the
worker error rate
• Membership Query (yes/no answer)
• Active Learning to carefully select
queries
• Dynamic Expressiveness of the
inference language
Costs
2/15

12. Scaling Wrapper Inference
Scaling the number of workers with Crowdsourcing platforms opens new
challenges:
Issues: Contributions:
Non-expert
workers
• Simple interactions to reduce the
worker error rate
• Membership Query (yes/no answer)
• Active Learning to carefully select
queries
• Dynamic Expressiveness of the
inference language
Costs
2/15
Quality
• Bayesian Model to evaluate the
expected wrapper quality
• Sampling algorithms

13. ALFRED
ALFRED is a wrapper inference system supervised by workers from a
crowdsourcing platform.
Input annotated page (page0):
3/15

14. ALFRED
ALFRED is a wrapper inference system supervised by workers from a
crowdsourcing platform.
r1 = /html/table/tr[1]/td/text()
r2 = //*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
r3 = //*[contains(.,”Director:”)]/../../tr[1]/td/text()
....
Inference
algorithm!
Input annotated page (page0):
3/15

15. ALFRED
ALFRED is a wrapper inference system supervised by workers from a
crowdsourcing platform.
r1 = /html/table/tr[1]/td/text()
r2 = //*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
r3 = //*[contains(.,”Director:”)]/../../tr[1]/td/text()
....
Inference
algorithm!
page0
r1
r2
r3
Spirited Away
Spirited Away
Spirited Away
Input annotated page (page0):
3/15

16. ALFRED
ALFRED is a wrapper inference system supervised by workers from a
crowdsourcing platform.
r1 = /html/table/tr[1]/td/text()
r2 = //*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
r3 = //*[contains(.,”Director:”)]/../../tr[1]/td/text()
....
Inference
algorithm!
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
Input annotated page (page0):
3/15

17. ALFRED
ALFRED is a wrapper inference system supervised by workers from a
crowdsourcing platform.
r1 = /html/table/tr[1]/td/text()
r2 = //*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
r3 = //*[contains(.,”Director:”)]/../../tr[1]/td/text()
....
Inference
algorithm!
Input annotated page (page0):
Is this title the correct one?
3/15

18. ALFRED
ALFRED is a wrapper inference system supervised by workers from a
crowdsourcing platform.
r1 = /html/table/tr[1]/td/text()
r2 = //*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
r3 = //*[contains(.,”Director:”)]/../../tr[1]/td/text()
....
Inference
algorithm!
DB#Wrapper!
r1 = /html/table/tr[1]/td/text()
Input annotated page (page0):
Is this title the correct one?
3/15

19. Membership Query
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
4/15
Yes !

20. Membership Query
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
• Rules compatible with the answer more
likely to be correct (Bayesian Model)
For each new answer
4/15
Yes !

21. Membership Query
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
• Rules compatible with the answer more
likely to be correct (Bayesian Model)
For each new answer
• If no rule is good enough:
• a new query is selected (Active Learning)
4/15
Yes !

22. Bayesian Model
Training sequence
= {“Spirited Away” , “-” , “9.3” }
Yes No No
5/15
Lk
Lk

23. Bayesian Model
Training sequence
= {“Spirited Away” , “-” , “9.3” }
Yes No No
5/15
Lk
Lk
a rule r is correct:
none of the candidate rules is correct:
Probability that:
P(r|Lk
)
P(R|Lk
)

24. Bayesian update:
Bayesian Model
Training sequence
= {“Spirited Away” , “-” , “9.3” }
Yes No No
5/15
Lk
Lk
a rule r is correct:
none of the candidate rules is correct:
Probability that:
P(r|Lk
)
P(R|Lk
)

25. Bayesian update:
Bayesian Model
Training sequence
= {“Spirited Away” , “-” , “9.3” }
Yes No No
5/15
Lk
Lk
a rule r is correct:
none of the candidate rules is correct:
Probability that:
P(r|Lk
)
P(R|Lk
)

26. Active Learning
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
ALFRED actively selects the queries;
a good policy saves money
6/15

27. Active Learning
• Random (baseline)
Values are randomly selected
• Entropy
Values are selected by maximizing the Entropy (most uncertain value)
• Greedy
Values are selected by minimizing the queries to confirm the most likely rule
• Lucky
Hybrid approach, it starts with an Entropy algorithm and then switch to Greedy to
confirm the best rule
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
ALFRED actively selects the queries;
a good policy saves money
6/15

28. Expressiveness
The candidate rules are generated observing the first annotated page
Should we use all the XPath expressiveness or just a fragment?
7/15
Expressiveness of the fragment Number of candidate rules

29. Expressiveness
Pool of candidate rules organized in fragments:
The candidate rules are generated observing the first annotated page
Should we use all the XPath expressiveness or just a fragment?
7/15
Expressiveness of the fragment Number of candidate rules

30. Expressiveness
Pool of candidate rules organized in fragments:
/html/table/tr[1]/td/text() Absolute Rules (complete path from root)
The candidate rules are generated observing the first annotated page
Should we use all the XPath expressiveness or just a fragment?
7/15
Expressiveness of the fragment Number of candidate rules

31. Expressiveness
Pool of candidate rules organized in fragments:
/html/table/tr[1]/td/text() Absolute Rules (complete path from root)
//*[contains(.,”Spirited Away”)]/text()
//*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
//*[contains(.,”Director:”)]/../../tr[1]/td/text()
Relative Rules (path from a textual node)
The candidate rules are generated observing the first annotated page
Should we use all the XPath expressiveness or just a fragment?
7/15
Expressiveness of the fragment Number of candidate rules

32. Expressiveness
Pool of candidate rules organized in fragments:
/html/table/tr[1]/td/text() Absolute Rules (complete path from root)
//*[contains(.,”Spirited Away”)]/text()
//*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
//*[contains(.,”Director:”)]/../../tr[1]/td/text()
Relative Rules (path from a textual node)
The candidate rules are generated observing the first annotated page
.... other XPaths
Should we use all the XPath expressiveness or just a fragment?
7/15
Expressiveness of the fragment Number of candidate rules

33. Expressiveness
/html/table/tr[1]/td/text()
//*[contains(.,”Spirited Away”)]/text()
//*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
//*[contains(.,”Director:”)]/../../tr[1]/td/text()
Correct (absolute) rule:
/html/table/tr[1]/td/text()
• The fragment is too expressive:
the correct rule can be generated
• But many MQ are needed to find it
8/15

34. Expressiveness
• The fragment is just expressive enough:
the correct rule can be generated.
• Few queries are needed to find it
/html/table/tr[1]/td/text()
/html/table/tr[1]/td/text()
//*[contains(.,”Spirited Away”)]/text()
//*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
//*[contains(.,”Director:”)]/../../tr[1]/td/text()
Correct (absolute) rule:
/html/table/tr[1]/td/text()
• The fragment is too expressive:
the correct rule can be generated
• But many MQ are needed to find it
8/15

35. Expressiveness
• The fragment is just expressive enough:
the correct rule can be generated.
• Few queries are needed to find it
/html/table/tr[1]/td/text()
/html/table/tr[1]/td/text()
//*[contains(.,”Spirited Away”)]/text()
//*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
//*[contains(.,”Director:”)]/../../tr[1]/td/text()
Correct (absolute) rule:
/html/table/tr[1]/td/text()
• The fragment is too expressive:
the correct rule can be generated
• But many MQ are needed to find it
8/15
State-of-the-art approaches fall in the first case !
They statically define the expressiveness of the XPath fragment

36. Expressiveness
• The fragment is just expressive enough:
the correct rule can be generated.
• Few queries are needed to find it
/html/table/tr[1]/td/text()
/html/table/tr[1]/td/text()
//*[contains(.,”Spirited Away”)]/text()
//*[contains(.,”Ratings:”)]/../../tr[1]/td/text()
//*[contains(.,”Director:”)]/../../tr[1]/td/text()
Correct (absolute) rule:
/html/table/tr[1]/td/text()
• The fragment is too expressive:
the correct rule can be generated
• But many MQ are needed to find it
8/15
State-of-the-art approaches fall in the first case !
They statically define the expressiveness of the XPath fragment

37. R0 : Absolute Rules
R1 : R0 + Relative Rules
.....
Expressiveness
5%
70%
25%
We defined simple XPath fragments.
Empirically observed: too expressive fragments are not actually needed.
9/15

38. Rules are organized in a Hierarchy of Fragments with increasing expressiveness
R0 : Absolute Rules
R1 : R0 + Relative Rules
.....
Expressiveness
5%
70%
25%
We defined simple XPath fragments.
Empirically observed: too expressive fragments are not actually needed.
9/15

39. Rules are organized in a Hierarchy of Fragments with increasing expressiveness
R0 : Absolute Rules
R1 : R0 + Relative Rules
.....
Inspired by Structural Risk Minimization (SRM)*:
a Machine Learning technique to address overfitting
*Details: Shawe-Taylor et all - IEEE Transactions on Information Theory, 44(5):1926–1940, 1998
Expressiveness
5%
70%
25%
We defined simple XPath fragments.
Empirically observed: too expressive fragments are not actually needed.
9/15

40. Dynamic Expressiveness
R0 : Absolute Rules
10/15

41. Dynamic Expressiveness
R0 : Absolute Rules
10/15
P(R|Lk
)
No solution?
> ?R

42. Dynamic Expressiveness
R0 : Absolute Rules
10/15
P(R|Lk
)
No solution?
> ?R
Expands the expressiveness
No

43. R1 : R0 + Relative Rules
Dynamic Expressiveness
R0 : Absolute Rules
10/15
P(R|Lk
)
No solution?
> ?R
Expands the expressiveness
No

44. .....
R1 : R0 + Relative Rules
Dynamic Expressiveness
R0 : Absolute Rules
10/15
P(R|Lk
)
No solution?
> ?R
Expands the expressiveness
No

45. .....
R1 : R0 + Relative Rules
Dynamic Expressiveness
R0 : Absolute Rules
10/15
P(r|Lk
)
Is r good enough?
> ?r
Expands the expressiveness
No

46. .....
R1 : R0 + Relative Rules
Yes
Terminates
Dynamic Expressiveness
R0 : Absolute Rules
10/15
P(r|Lk
)
Is r good enough?
> ?r
Expands the expressiveness
No

47. Results
Site Entity |Pages|
www.imdb.com Actor 500k
www.imdb.com Movies 500k
www.allmusic.com Band 500k
www.allmusic.com Albums 500k
www.nasdaq.com Stock Quotes 7k
Dataset: 40 attributes
Measures:
• Costs - #MQ
• Quality - Precision and Recall
11/15

48. Results: Dynamic Expressiveness
Strategy #MQ (SRM off) #MQ (SRM on) % MQ saved P (SRM on) R (SRM on)
RANDOM 379 190 50% 0,998 0,977
GREEDY 398 169 58% 0,998 0,983
LUCKY 196 132 33% 0,996 0,995
ENTROPY 205 116 44% 0,998 0,99
12/15

49. Results: Dynamic Expressiveness
Strategy #MQ (SRM off) #MQ (SRM on) % MQ saved P (SRM on) R (SRM on)
RANDOM 379 190 50% 0,998 0,977
GREEDY 398 169 58% 0,998 0,983
LUCKY 196 132 33% 0,996 0,995
ENTROPY 205 116 44% 0,998 0,99
Dynamic Expressiveness saves a lot of queries
12/15

50. Results: Dynamic Expressiveness
Strategy #MQ (SRM off) #MQ (SRM on) % MQ saved P (SRM on) R (SRM on)
RANDOM 379 190 50% 0,998 0,977
GREEDY 398 169 58% 0,998 0,983
LUCKY 196 132 33% 0,996 0,995
ENTROPY 205 116 44% 0,998 0,99
Dynamic Expressiveness saves a lot of queries
Small quality loss:
The expressiveness is not expanded when it is needed
12/15

51. Results: Dynamic Expressiveness
Static Expressiveness Dynamic Expressiveness
# candidate rules # candidate rules
13/15

52. Results: Dynamic Expressiveness
Static Expressiveness Dynamic Expressiveness
“Simple” attributes: complex algorithms are not needed
# candidate rules # candidate rules
13/15

53. Results: Dynamic Expressiveness
Static Expressiveness Dynamic Expressiveness
“Simple” attributes: complex algorithms are not needed
“Complex” attributes: Entropy, Lucky and Dynamic Expressiveness saves
a lot of queries
# candidate rules # candidate rules
13/15

54. Future development
Noisy Crowds: workers mistakes vs task redundancy*
How to evaluate the accuracy of the worker?
Another query or another worker?
Same learning framework, different problems: NLP, Crawling
14/15
*Demo
Title: ALFRED: Crowd Assisted Data Extraction
When: Tomorrow 17h
Where: Imperial Room

55. Thank you for the attention !!
15/15

56. 15/15
Redundancy
0
0,5
1
0 1 2 3 4
P(r1)
P(r2)
P(r3)
# MQ
0
0,5
1
0 1 2 3 4
P(r1)
P(r2)
P(r3)
Not Accurate Worker
# MQ
0
0,5
1
0 1 2 3 4
P(r1)
P(r2)
P(r3)
# MQ
Many Workers
Accurate Worker

57. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....

58. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....

59. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....
Inference
algorithm!

60. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....
Inference
algorithm!

61. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....
Inference
algorithm!

62. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....
Inference
algorithm!

63. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....
Wrapper!
Inference
algorithm!

64. ... selecting the right sample set is crucial
Sampling & Quality
2M pages from IMDB, we have to work with a sample set but ....
Wrapper!
Inference
algorithm!
DB#
... Not all pages look like the pages about famous movies

65. Sampling & Quality
page0
r1
r2
r3
Spirited Away
Spirited Away
Spirited Away
r1 = r2 = r3

66. Sampling & Quality
page0
r1
r2
r3
Spirited Away
Spirited Away
Spirited Away
r1 = r2 = r3
page0 page1
r1
r2
r3
Spirited Away City of God
Spirited Away -
Spirited Away City of God
r1 = r3 != r2

67. Sampling & Quality
page0
r1
r2
r3
Spirited Away
Spirited Away
Spirited Away
r1 = r2 = r3
page0 page1
r1
r2
r3
Spirited Away City of God
Spirited Away -
Spirited Away City of God
r1 = r3 != r2
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
r1 != r3 != r2

68. Sampling & Quality
page0
r1
r2
r3
Spirited Away
Spirited Away
Spirited Away
r1 = r2 = r3
page0 page1
r1
r2
r3
Spirited Away City of God
Spirited Away -
Spirited Away City of God
r1 = r3 != r2
page0 page1 page2
r1
r2
r3
Spirited Away City of God Howl’s Moving Castle
Spirited Away - 9.3
Spirited Away City of God null
r1 != r3 != r2
Pages make apparent the
differences among the rules
Find a small set that makes apparent
the same differences observed in the
whole set of pages*

69. Sampling & Quality
The problem.
Find the smallest set that makes apparent the differences among the rules:
(e.g., 100 pages that make apparent the same differences that we would observe in 2M pages).
It is a NP-Hard problem !! Reduction to SET-Cover problem:
Find the smallest set of pages that cover all the group of rules (group = equivalent rules).
The smallest set is not needed:
A greedy algorithm O(|Pages|) in time and O(1) in space works very well in practice.

70. XPath rules
For every page p:
if (p makes apparent new differences)
representative pages += p
An offline algorithm that can be easily parallelized
Sampling & Quality

71. Results: Sampling
Three sample sets:
• Biased
Pages collected by crawling the website
• Random
Pages randomly picked from the whole set of pages
• Representative
Pages collected by our sampling algorithm

72. Results: Sampling
Entity Sampling |Pages| P R
Movies
Biased 250 0.98 0.71
Movies Random 250 0.99 0.99Movies
Representative 42 1.00 1.00
Actors
Biased 250 1.00 1.00
Actors Random 250 1.00 0.96Actors
Representative 30 1.00 1.00
Stocks
Biased 86 1.00 0.98
Stocks Random 86 1.00 0.99Stocks
Representative 15 1.00 1.00
Albums
Biased 258 1.00 0.99
Albums Random 258 1.00 1.00Albums
Representative 59 1.00 1.00
Bands
Biased 289 1.00 0.68
Bands Random 289 1.00 1.00Bands
Representative 36 1.00 1.00

73. Results: Sampling
Entity Sampling |Pages| P R
Movies
Biased 250 0.98 0.71
Movies Random 250 0.99 0.99Movies
Representative 42 1.00 1.00
Actors
Biased 250 1.00 1.00
Actors Random 250 1.00 0.96Actors
Representative 30 1.00 1.00
Stocks
Biased 86 1.00 0.98
Stocks Random 86 1.00 0.99Stocks
Representative 15 1.00 1.00
Albums
Biased 258 1.00 0.99
Albums Random 258 1.00 1.00Albums
Representative 59 1.00 1.00
Bands
Biased 289 1.00 0.68
Bands Random 289 1.00 1.00Bands
Representative 36 1.00 1.00
Representative perfect

74. Results: Sampling
Entity Sampling |Pages| P R
Movies
Biased 250 0.98 0.71
Movies Random 250 0.99 0.99Movies
Representative 42 1.00 1.00
Actors
Biased 250 1.00 1.00
Actors Random 250 1.00 0.96Actors
Representative 30 1.00 1.00
Stocks
Biased 86 1.00 0.98
Stocks Random 86 1.00 0.99Stocks
Representative 15 1.00 1.00
Albums
Biased 258 1.00 0.99
Albums Random 258 1.00 1.00Albums
Representative 59 1.00 1.00
Bands
Biased 289 1.00 0.68
Bands Random 289 1.00 1.00Bands
Representative 36 1.00 1.00

75. Results: Sampling
Entity Sampling |Pages| P R
Movies
Biased 250 0.98 0.71
Movies Random 250 0.99 0.99Movies
Representative 42 1.00 1.00
Actors
Biased 250 1.00 1.00
Actors Random 250 1.00 0.96Actors
Representative 30 1.00 1.00
Stocks
Biased 86 1.00 0.98
Stocks Random 86 1.00 0.99Stocks
Representative 15 1.00 1.00
Albums
Biased 258 1.00 0.99
Albums Random 258 1.00 1.00Albums
Representative 59 1.00 1.00
Bands
Biased 289 1.00 0.68
Bands Random 289 1.00 1.00Bands
Representative 36 1.00 1.00
Biased: recall loss

76. Results: Sampling
Entity Sampling |Pages| P R
Movies
Biased 250 0.98 0.71
Movies Random 250 0.99 0.99Movies
Representative 42 1.00 1.00
Actors
Biased 250 1.00 1.00
Actors Random 250 1.00 0.96Actors
Representative 30 1.00 1.00
Stocks
Biased 86 1.00 0.98
Stocks Random 86 1.00 0.99Stocks
Representative 15 1.00 1.00
Albums
Biased 258 1.00 0.99
Albums Random 258 1.00 1.00Albums
Representative 59 1.00 1.00
Bands
Biased 289 1.00 0.68
Bands Random 289 1.00 1.00Bands
Representative 36 1.00 1.00

77. Results: Sampling
Entity Sampling |Pages| P R
Movies
Biased 250 0.98 0.71
Movies Random 250 0.99 0.99Movies
Representative 42 1.00 1.00
Actors
Biased 250 1.00 1.00
Actors Random 250 1.00 0.96Actors
Representative 30 1.00 1.00
Stocks
Biased 86 1.00 0.98
Stocks Random 86 1.00 0.99Stocks
Representative 15 1.00 1.00
Albums
Biased 258 1.00 0.99
Albums Random 258 1.00 1.00Albums
Representative 59 1.00 1.00
Bands
Biased 289 1.00 0.68
Bands Random 289 1.00 1.00Bands
Representative 36 1.00 1.00
Random:
better than biased

78. State of Art
• 2006 - Interactive wrapper generation with minimal user effort.
U. Irmik et al. WWW
• 2006 - Active learning with multiple views.
I. Muslea et al. JAIR
Supervised
Wrapper Induction

79. State of Art
• 2008 - Wrapper inference for ambiguous web pages.
C. Valter and P. Merialdo JAAI
• 2005 - Web Data Extraction Based on Partial Tree Alignment
Yanhong Zhai WWW.
Unsupervised
Wrapper Induction

80. State of Art
• 2012 - D.I.A.D.E.M.
J. Furche and G. Gottlob WWW
• 2011 - Automatic wrappers for large scale web extraction.
N.N. Dalvi et al. VLDB.
Automatic Annotators