Is Fidel Castro Really an American President? On Set Expansion An - - PowerPoint PPT Presentation

▶
is fidel castro really an american president
SMART_READER_LITE
LIVE PREVIEW

Is Fidel Castro Really an American President? On Set Expansion An - - PowerPoint PPT Presentation

Feb 09, 2024 118 likes •1.41k views

Is Fidel Castro Really an American President? On Set Expansion An Example Google Sets Google Sets Google Sets Google Sets Notion Notion Seeds: Barack Obama, Bill Clinton, George Bush Notion Seeds: Barack Obama, Bill Clinton,

slide-1
SLIDE 1

Is Fidel Castro Really an American President?

On Set Expansion

slide-2
SLIDE 2

An Example

slide-3
SLIDE 3

Google Sets

slide-4
SLIDE 4

Google Sets

slide-5
SLIDE 5

Google Sets

slide-6
SLIDE 6

Google Sets

slide-7
SLIDE 7

Notion

slide-8
SLIDE 8

Notion

  • Seeds: Barack Obama, Bill Clinton,

George Bush

slide-9
SLIDE 9

Notion

  • Seeds: Barack Obama, Bill Clinton,

George Bush

  • Target Set: US-American presidents
slide-10
SLIDE 10

Notion

  • Seeds: Barack Obama, Bill Clinton,

George Bush

  • Target Set: US-American presidents
  • Answer to query: more probable

elements of the target set

slide-11
SLIDE 11

Notion

  • Seeds: Barack Obama, Bill Clinton,

George Bush

  • Target Set: US-American presidents
  • Answer to query: more probable

elements of the target set

  • John F. Kennedy
slide-12
SLIDE 12

Notion

  • Seeds: Barack Obama, Bill Clinton,

George Bush

  • Target Set: US-American presidents
  • Answer to query: more probable

elements of the target set

  • John F. Kennedy
  • George Washington, etc.
slide-13
SLIDE 13

SEAL

slide-14
SLIDE 14

The SEAL System

slide-15
SLIDE 15

The SEAL System

Seeds

slide-16
SLIDE 16

The SEAL System

Seeds Web pages

Fetcher

Google

slide-17
SLIDE 17

The SEAL System

Seeds Web pages

Fetcher

Google

Extractor

Mentions Wrapper

slide-18
SLIDE 18

The SEAL System

Seeds Web pages

Fetcher

Google

Extractor

Mentions Wrapper

Ranker

Suggestions Graph

slide-19
SLIDE 19

The SEAL System

Seeds Web pages

Fetcher

Google

Extractor

Mentions Wrapper

Ranker

Suggestions Graph

slide-20
SLIDE 20

Fetcher

slide-21
SLIDE 21

Fetcher

slide-22
SLIDE 22
  • Simple Google search

Fetcher

slide-23
SLIDE 23
  • Simple Google search
  • Query: Concatenation of all seeds

Fetcher

slide-24
SLIDE 24
  • Simple Google search
  • Query: Concatenation of all seeds
  • Crawl top results

Fetcher

slide-25
SLIDE 25

The SEAL System

Seeds Web pages

Fetcher

Google

Extractor

Mentions Wrapper

Ranker

Suggestions Graph

slide-26
SLIDE 26

Extractor

slide-27
SLIDE 27

Extractor

  • Idea:
slide-28
SLIDE 28

Extractor

  • Idea:
  • Find common contexts of all seeds
slide-29
SLIDE 29

Extractor

  • Idea:
  • Find common contexts of all seeds
  • Wrapper
slide-30
SLIDE 30

Extractor

  • Idea:
  • Find common contexts of all seeds
  • Wrapper
  • Derive new entities
slide-31
SLIDE 31

Extractor

  • Idea:
  • Find common contexts of all seeds
  • Wrapper
  • Derive new entities

<ul> <li>Obama</li> <li>Bush</li> <li>Kennedy</li> </ul>

slide-32
SLIDE 32

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

slide-33
SLIDE 33

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

slide-34
SLIDE 34

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds Right Contexts

slide-35
SLIDE 35

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • xxaEsd...

Right Contexts

slide-36
SLIDE 36

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • xxaEsd...
  • xkhSlp...

Right Contexts

slide-37
SLIDE 37

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • xxaEsd...
  • xkhSlp...
  • gzPHMA...

Right Contexts

slide-38
SLIDE 38

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • xxaEsd...
  • xkhSlp...
  • gzPHMA...
  • pfblRo...

Right Contexts

slide-39
SLIDE 39

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil Document

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • xxaEsd...
  • xkhSlp...
  • gzPHMA...
  • pfblRo...
  • gzUmXu...

Right Contexts

slide-40
SLIDE 40

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil

Obama Clinton 1 Obama 2 Clinton 3 Clinton 4

  • xxaEsd...
  • xkhSlp...
  • gzPHMA...
  • pfblRo...
  • gzUmXu...

Document Seeds Right Contexts

slide-41
SLIDE 41
  • Practical Algorithm

To Retrieve Information Coded In Alphanumeric

Patricia Trie

slide-42
SLIDE 42

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil

Obama 0 Clinton 1 Obama 2 Clinton 3 Clinton 4

“” {0,1,2,3,4} “pfblRo...” {3} “x” {0,1} “gz” {2,4}

  • xxaEsd...
  • xkhSlp...
  • gzPHMA...
  • pfblRo...
  • gzUmXu...

“xaEsd...” {0} “khSlp...” {1} “PHMA...” {2} “UmXu...” {4}

Document Seeds Right Contexts Trie Tright

slide-43
SLIDE 43

“” {0,1,2,3,4} “trp” {0,1} “Ag” {2,3,4} “ASxy...” {1} “f” {2,4} “Lloc...” {3} “” {0,1,2,3,4} “pfblRo...” {3} “x” {0,1} “gz” {2,4} “xaEsd...” {0} “khSlp...” {1} “PHMA...” {2} “UmXu...” {4}

Tright Tleft

“Upl...” {2} “TiL...” {4}

Obama 0 Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

slide-44
SLIDE 44

“” {0,1,2,3,4} “trp” {0,1} “Ag” {2,3,4} “ASxy...” {1} “f” {2,4} “Lloc...” {3} “” {0,1,2,3,4} “pfblRo...” {3} “x” {0,1} “gz” {2,4} “xaEsd...” {0} “khSlp...” {1} “PHMA...” {2} “UmXu...” {4}

Tright Tleft

“Upl...” {2} “TiL...” {4}

Obama 0 Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • 1. TopNodes

Algorithm

slide-45
SLIDE 45

“” {0,1,2,3,4} “trp” {0,1} “Ag” {2,3,4} “ASxy...” {1} “f” {2,4} “Lloc...” {3} “” {0,1,2,3,4} “pfblRo...” {3} “x” {0,1} “gz” {2,4} “xaEsd...” {0} “khSlp...” {1} “PHMA...” {2} “UmXu...” {4}

Tright Tleft

“Upl...” {2} “TiL...” {4}

Obama 0 Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • 1. TopNodes
  • 2. Match

Algorithm

slide-46
SLIDE 46

“” {0,1,2,3,4} “trp” {0,1} “Ag” {2,3,4} “ASxy...” {1} “f” {2,4} “Lloc...” {3} “” {0,1,2,3,4} “pfblRo...” {3} “x” {0,1} “gz” {2,4} “xaEsd...” {0} “khSlp...” {1} “PHMA...” {2} “UmXu...” {4}

Tright Tleft

“Upl...” {2} “TiL...” {4}

Obama 0 Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

  • 1. TopNodes
  • 2. Match
  • 3. Match back

Algorithm

slide-47
SLIDE 47

“” {0,1,2,3,4} “trp” {0,1} “Ag” {2,3,4} “ASxy...” {1} “f” {2,4} “Lloc...” {3} “” {0,1,2,3,4} “pfblRo...” {3} “x” {0,1} “gz” {2,4} “xaEsd...” {0} “khSlp...” {1} “PHMA...” {2} “UmXu...” {4}

Tright Tleft

“Upl...” {2} “TiL...” {4}

Obama 0 Clinton 1 Obama 2 Clinton 3 Clinton 4

Seeds

slide-48
SLIDE 48

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil

Wrapper: prt[...]x Content:

  • bama, clinton

Wrapper: fgA[...]gz Content:

  • bama, clinton

Document

slide-49
SLIDE 49

prtoBamAxxaEsdSlkprtKenNed yxSAprtCLinTOnxkhSlpUfgAob AMagzPHMAcolLgAcLIntOnpfb lRoiusWgoprtcAstrOxkLiTfgAClI nTongzUmXuSYfgAkEnneDygzil

Wrapper: prt[...]x Content:

  • bama, clinton, kennedy, castro

Wrapper: fgA[...]gz Content:

  • bama, clinton, kennedy

Document

slide-50
SLIDE 50

Characters vs Tags

slide-51
SLIDE 51

Characters vs Tags

<ul> <li>Obama</li> <li>Bush</li> <li>Kennedy</li> </ul>

slide-52
SLIDE 52
  • Exploit (HTML) tags

Characters vs Tags

<ul> <li>Obama</li> <li>Bush</li> <li>Kennedy</li> </ul>

slide-53
SLIDE 53
  • Exploit (HTML) tags
  • Good idea?

Characters vs Tags

<ul> <li>Obama</li> <li>Bush</li> <li>Kennedy</li> </ul>

slide-54
SLIDE 54
  • Exploit (HTML) tags
  • Good idea?
  • Apparently not

Characters vs Tags

<ul> <li>Obama</li> <li>Bush</li> <li>Kennedy</li> </ul>

slide-55
SLIDE 55

Characters vs Tags

slide-56
SLIDE 56

Characters vs Tags

  • No parser
slide-57
SLIDE 57

Characters vs Tags

  • No parser
  • Language-independent
slide-58
SLIDE 58

Characters vs Tags

  • No parser
  • Language-independent
  • ⠟⠁⠃⠹⠕⠊⠟⠭⠛⠑⠹
slide-59
SLIDE 59

Characters vs Tags

  • No parser
  • Language-independent
  • ⠟⠁⠃⠹⠕⠊⠟⠭⠛⠑⠹
  • Other meta language (e.g. TeX)
slide-60
SLIDE 60

Characters vs Tags

  • No parser
  • Language-independent
  • ⠟⠁⠃⠹⠕⠊⠟⠭⠛⠑⠹
  • Other meta language (e.g. TeX)
  • Seeds in unusual places
slide-61
SLIDE 61

Characters vs Tags

  • No parser
  • Language-independent
  • ⠟⠁⠃⠹⠕⠊⠟⠭⠛⠑⠹
  • Other meta language (e.g. TeX)
  • Seeds in unusual places
  • <span class=”obama”>
slide-62
SLIDE 62

Characters vs Tags

  • No parser
  • Language-independent
  • ⠟⠁⠃⠹⠕⠊⠟⠭⠛⠑⠹
  • Other meta language (e.g. TeX)
  • Seeds in unusual places
  • <span class=”obama”>
  • More restrictions ⇒ lower performance
slide-63
SLIDE 63

The SEAL System

Seeds Web pages

Fetcher

Google

Extractor

Mentions Wrapper

Ranker

Suggestions Graph

slide-64
SLIDE 64

Ranker

slide-65
SLIDE 65

Ranker

  • Problem: Noise
slide-66
SLIDE 66

Ranker

  • Problem: Noise
  • Solution: Similarity measure between

seeds and mentions

slide-67
SLIDE 67

Ranker

  • Problem: Noise
  • Solution: Similarity measure between

seeds and mentions

  • Ranked output
slide-68
SLIDE 68

Ranker

  • Problem: Noise
  • Solution: Similarity measure between

seeds and mentions

  • Ranked output
  • Understand relation
slide-69
SLIDE 69

Analysis

slide-70
SLIDE 70

Analysis

slide-71
SLIDE 71

Analysis

slide-72
SLIDE 72

Analysis

slide-73
SLIDE 73

Analysis

slide-74
SLIDE 74

Analysis

  • 4 source types:
  • document
  • seed
  • wrapper
  • mention
slide-75
SLIDE 75

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

slide-76
SLIDE 76

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

Graph-walk (Page rank)

slide-77
SLIDE 77

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

Graph-walk (Page rank) P(find|doc) = 0.5

slide-78
SLIDE 78

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

Graph-walk (Page rank) P(find|doc) = 0.5 P(derive|doc) = 0.5

slide-79
SLIDE 79

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

Graph-walk (Page rank) P(find|doc) = 0.5 P(derive|doc) = 0.5 P(seeds|doc,find) = 1

slide-80
SLIDE 80

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

Graph-walk (Page rank) P(find|doc) = 0.5 P(derive|doc) = 0.5 P(seeds|doc,find) = 1 P(prt..x|doc,derive) = 0.5

slide-81
SLIDE 81

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

Graph-walk (Page rank) P(find|doc) = 0.5 P(derive|doc) = 0.5 P(seeds|doc,find) = 1 P(prt..x|doc,derive) = 0.5 P(fgA..gz|doc,derive) = 0.5

slide-82
SLIDE 82

doc seeds

prt..x Castro JFK fgA..gz

1/2 1/4 1/4 extract extract extract

Graph-walk (Page rank) P(find|doc) = 0.5 P(derive|doc) = 0.5 P(seeds|doc,find) = 1 P(prt..x|doc,derive) = 0.5 P(fgA..gz|doc,derive) = 0.5

slide-83
SLIDE 83

doc seeds

prt..x Castro JFK fgA..gz

find derive 1/2 extract 1/4 1/4

Graph-walk (Page rank) Transitions in both ways

slide-84
SLIDE 84

s d w1 w2 m1 m2 s d w1 w2 m1 m2

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-85
SLIDE 85

x y

s d w1 w2 m1 m2 s d w1 w2 m1 m2

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-86
SLIDE 86

x y

s d w1 w2 m1 m2 s d w1 w2 m1 m2

Transition Matrix (x,y) = P(x→y)

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-87
SLIDE 87

s d w1 w2 m1 m2 s d w1 w2 m1 m2

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-88
SLIDE 88

s d w1 w2 m1 m2 s d 1 w1 w2 m1 m2

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-89
SLIDE 89

s d w1 w2 m1 m2 s d 1 w1 w2 m1 m2

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-90
SLIDE 90

s d w1 w2 m1 m2 s ½ d 1 w1 ¼ w2 ¼ m1 m2

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-91
SLIDE 91

s d w1 w2 m1 m2 s ½ d 1 ½ ½ w1 ¼ ½ 1 w2 ¼ ½ m1 ¼ ½ m2 ¼

Transition Matrix

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-92
SLIDE 92

doc seeds

prt..x Castro JFK fgA..gz

find derive derive extract extract extract

s d w1 w2 m1 m2 s ½ d 1 ½ ½ w1 ¼ ½ 1 w2 ¼ ½ m1 ¼ ½ m2 ¼

Transition Matrix

with lazyness factor λ=0.01

slide-93
SLIDE 93

s d w1 w2 m1 m2

State Vector

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-94
SLIDE 94

s 1 d w1 w2 m1 m2

State Vector

d s

w2 m2 m1 w1

find derive derive extract extract extract

slide-95
SLIDE 95

Transition Matrix and State Vector

slide-96
SLIDE 96

Transition Matrix and State Vector · =

slide-97
SLIDE 97

Transition Matrix and State Vector · =

slide-98
SLIDE 98

Iterated Multiplication · =

slide-99
SLIDE 99

Iterated Multiplication · =

1000x

slide-100
SLIDE 100

Iterated Multiplication · =

1000x

slide-101
SLIDE 101

Iterated Multiplication · =

d s

w2 m2 m1 w1

find derive derive extract extract extract

1000x

slide-102
SLIDE 102

Evaluation

slide-103
SLIDE 103
  • Tested English, Chinese, Japanese

Evaluation

slide-104
SLIDE 104
  • Tested English, Chinese, Japanese
  • 36 datasets, 18 classes

Evaluation

slide-105
SLIDE 105
  • Tested English, Chinese, Japanese
  • 36 datasets, 18 classes
  • 9 classes for all 3 languages (e.g. US

presidents)

Evaluation

slide-106
SLIDE 106
  • Tested English, Chinese, Japanese
  • 36 datasets, 18 classes
  • 9 classes for all 3 languages (e.g. US

presidents)

  • 9 language-specific classes (e.g.

Chinese dynasties)

Evaluation

slide-107
SLIDE 107
  • Tested English, Chinese, Japanese
  • 36 datasets, 18 classes
  • 9 classes for all 3 languages (e.g. US

presidents)

  • 9 language-specific classes (e.g.

Chinese dynasties)

  • 5 runs/dataset, 3 random seeds

Evaluation

slide-108
SLIDE 108

MAP >90%

slide-109
SLIDE 109

US Presidents cont.

slide-110
SLIDE 110

US Presidents cont.

  • rcwang.com/seal
slide-111
SLIDE 111

US Presidents cont.

  • rcwang.com/seal
  • boowa.com
slide-112
SLIDE 112

US Presidents cont.

  • rcwang.com/seal
  • boowa.com
  • Still strange results
slide-113
SLIDE 113

US Presidents cont.

  • rcwang.com/seal
  • boowa.com
  • Still strange results
  • Google results
slide-114
SLIDE 114

US Presidents cont.

  • rcwang.com/seal
  • boowa.com
  • Still strange results
  • Google results
  • Latest news
slide-115
SLIDE 115

US Presidents cont.

  • rcwang.com/seal
  • boowa.com
  • Still strange results
  • Google results
  • Latest news
slide-116
SLIDE 116

US Presidents cont.

  • rcwang.com/seal
  • boowa.com
  • Still strange results
  • Google results
  • Latest news
slide-117
SLIDE 117

Bootstrapping

slide-118
SLIDE 118

stats = {}, used = input, rslt = {} for i = 1 to M do m = min(3,|used|) seeds = selectm(used) ∪ top(rslt) stats = expand(seeds) rslt = rank(stats) used = used ∪ seeds rof

slide-119
SLIDE 119

Binary Relations

slide-120
SLIDE 120

More Wrappers

slide-121
SLIDE 121

More Wrappers

  • New: Middle context
slide-122
SLIDE 122

More Wrappers

  • New: Middle context
  • Slight adjustments to the old algorithm
slide-123
SLIDE 123

More Wrappers

  • New: Middle context
  • Slight adjustments to the old algorithm
  • “Mayor of...”
slide-124
SLIDE 124

More Wrappers

  • New: Middle context
  • Slight adjustments to the old algorithm
  • “Mayor of...”
  • “Duet with...”, etc.
slide-125
SLIDE 125

sadhSAbcGsadutsobAMawrKu SAjkLsdFautsmeRKelwrKgErmA nyjkLsdfkuxeSAmcvDkBSs

slide-126
SLIDE 126

sadhSAbcGsadutsobAMawrKu SAjkLsdFautsmeRKelwrKgErmA nyjkLsdfkuxeSAmcvDkBSs

slide-127
SLIDE 127

Much research going on

Not sure if that fits...

slide-128
SLIDE 128

Thank you

Seeds Web pages

Fetcher

Google

Extractor

Mentions

Wrapper

Ranker

Suggestions

Graph

  • Bootstrapping
  • Binary relations