SlideShare a Scribd company logo
Detecting Emergent Intersectional Biases: 
Contextualized Word Embeddings Contain a Distribution of
Human-like Biases
Wei Guo Aylin Caliskan
August 1, 2021
George Washington University
1
Bias in NLP is everywhere!
• Bias in NLP perpetuates bias in society
• Incomprehensive measurement
• Cannot automatically identify bias
2
Intersectional bias is concerning!
3
Intersectional bias is concerning!
• Incomplete measurement
of social biases
• Unique experiences of
discrimination in ML system
3
Methods: implicit cognition → natural language → computer vision
Implicit Association Test (IAT)
• Tests for differential association of
two concepts
• Easier to categorize
stereotype-congruent pairs
• Harder to categorize
stereotype-incongruent pairs
• Effect d = difference in reaction time Weapon IAT (implicit.harvard.edu)
4
Methods: implicit cognition → natural language → computer vision
Implicit Association Test (IAT)
• Tests for differential association of
two concepts
• Easier to categorize
stereotype-congruent pairs
• Harder to categorize
stereotype-incongruent pairs
• Effect d = difference in reaction time
Weapon IAT (implicit.harvard.edu)
4
Methods: implicit cognition → natural language → computer vision
Implicit Association Test (IAT)
• Tests for differential association of
two concepts
• Easier to categorize
stereotype-congruent pairs
• Harder to categorize
stereotype-incongruent pairs
• Effect d = difference in reaction time
Weapon IAT (implicit.harvard.edu)
4
Methods: implicit cognition → natural language → computer vision
Implicit Association Test (IAT)
• Tests for differential association of
two concepts
• Easier to categorize
stereotype-congruent pairs
• Harder to categorize
stereotype-incongruent pairs
• Effect d = difference in reaction time
4
Methods: implicit cognition → static embeddings → contextualized embeddings
Word Embedding Association Test Implicit Association Test
5
Methods: implicit cognition → static embeddings → contextualized embeddings
Word Embedding Association Test
5
Methods: implicit cognition → static embeddings → contextualized embeddings
man
[
feature1 feature2 . . . featured
]
father
[
feature1 feature2 . . . featured
]
.
.
.
woman
[
feature1 feature2 . . . featured
]
mother
[
feature1 feature2 . . . featured
]
.
.
.
science
[
feature1 feature2 . . . featured
]
math
[
feature1 feature2 . . . featured
]
.
.
.
liberal arts
[
feature1 feature2 . . . featured
]
music
[
feature1 feature2 . . . featured
]
.
.
.
5
Methods: implicit cognition → static embeddings → contextualized embeddings
man
[
feature1 feature2 . . . featured
]
father
[
feature1 feature2 . . . featured
]
.
.
.
woman
[
feature1 feature2 . . . featured
]
mother
[
feature1 feature2 . . . featured
]
.
.
.
science
[
feature1 feature2 . . . featured
]
math
[
feature1 feature2 . . . featured
]
.
.
.
liberal arts
[
feature1 feature2 . . . featured
]
music
[
feature1 feature2 . . . featured
]
.
.
.
Word Embedding Association Test
(WEAT)
s(w, A, B) = meana∈A cos(w, a)−meanb∈B cos(w, b)
s(X, Y, A, B) =
∑
x∈X
s(x, A, B) −
∑
y∈Y
s(y, A, B)
5
Methods: implicit cognition → static embeddings → contextualized embeddings
Implicit Association Test
Word Embedding Factual Association
Test (WEFAT)
s(w, A, B) =
meana∈As(⃗
w,⃗
a) − meanb∈Bs(⃗
w,⃗
b)
stdx∈A∪Bs(⃗
w, x)
6
Intersectional Bias Detection (IBD)
s(w, A, B) =
meana∈As(⃗
w,⃗
a) − meanb∈Bs(⃗
w,⃗
b)
stdx∈A∪Bs(⃗
w, x)
7
Intersectional Bias Detection (IBD)
s(w, A, B) =
meana∈As(⃗
w,⃗
a) − meanb∈Bs(⃗
w,⃗
b)
stdx∈A∪Bs(⃗
w, x)
Detecting intersectional biases
associated with members of multiple minority groups.
7
Emergent Intersectional Bias Detection (EIBD)
8
Emergent Intersectional Bias Detection (EIBD)
Intersectional biases - Attributes highly associated with single social category =
Remaining set is the emergent intersectional biases
Detecting unique emergent intersectional biases that do not overlap with the
biases of their constituent minority identities. 
8
Evaluation of IBD
Detection accuracy > 80% accuracy, where random chance < 15%
Validation set for intersectional biases from Ghavami Peplau, 2013
9
Methods: implicit cognition → static embeddings → contextualized embeddings
Extract the sentence containing the words X, Y, A, B
Contextualized Embedding Assoiciation Test (CEAT)
10
Methods: implicit cognition → static embeddings → contextualized embeddings
Generates the contextualized embeddings
10
Methods: implicit cognition → static embeddings → contextualized embeddings
Calculate the effect size of bias based on WEAT
10
Contextualized Embedding Association Test (CEAT)
Generates the distribution of effect magnitudes of biases
Calculate Combined Effect Size
CES(X, Y, A, B) =
∑N
i=1 viESi
∑N
i=1 vi
10
Evaluation of CEAT
Contextualized
embeddings from Corpus of
• Widely shared biases
• Flowers/insects
• Musical instru-
ments/weapons
• Social group biases
• Gender
• Race
• Intersectionality
• ...
11
Evaluation of CEAT
• Intersectional biases have
high magnitude.
• Biased: ELMo > BERT >
GPT > GPT-2
• The overall magnitude of
bias negatively correlates
with the level of
contextualization in the
language model.
11
Questions?
weiguo@gwu.edu
github.com/weiguowilliam/CEAT
paper code
Acknowledgements
my co-author Aylin Caliskan & many reviewers
11

More Related Content

What's hot

Configure iis to access your website using an ip address
Configure iis to access your website using an ip addressConfigure iis to access your website using an ip address
Configure iis to access your website using an ip address
birhanu atnafu
 
A criança interior ferida
A criança interior feridaA criança interior ferida
A criança interior ferida
Cinara Aline
 
Os Frutos do Espírito
Os Frutos do EspíritoOs Frutos do Espírito
Os Frutos do Espírito
Freekidstories
 
Caderno formatura
Caderno formaturaCaderno formatura
Caderno formatura
muni2014
 
Reta vaca
Reta vacaReta vaca
Reta vaca
Guto Gomes Gomes
 
Auto estima 1a aula
Auto estima 1a aulaAuto estima 1a aula
Auto estima 1a aula
Atividades Diversas Cláudia
 
Como evangelizar crianças
Como evangelizar criançasComo evangelizar crianças
Como evangelizar crianças
Edleusa Silva
 
Terço de são bento
Terço de são bentoTerço de são bento
Terço de são bento
Ticiano Raphael de Mattos
 

What's hot (8)

Configure iis to access your website using an ip address
Configure iis to access your website using an ip addressConfigure iis to access your website using an ip address
Configure iis to access your website using an ip address
 
A criança interior ferida
A criança interior feridaA criança interior ferida
A criança interior ferida
 
Os Frutos do Espírito
Os Frutos do EspíritoOs Frutos do Espírito
Os Frutos do Espírito
 
Caderno formatura
Caderno formaturaCaderno formatura
Caderno formatura
 
Reta vaca
Reta vacaReta vaca
Reta vaca
 
Auto estima 1a aula
Auto estima 1a aulaAuto estima 1a aula
Auto estima 1a aula
 
Como evangelizar crianças
Como evangelizar criançasComo evangelizar crianças
Como evangelizar crianças
 
Terço de são bento
Terço de são bentoTerço de são bento
Terço de são bento
 

Similar to Detecting emergent intersectional biases: Contextualized word embeddings contain a distribution of human-like biases

From Human Intelligence to Machine Intelligence
From Human Intelligence to Machine IntelligenceFrom Human Intelligence to Machine Intelligence
From Human Intelligence to Machine Intelligence
NUS-ISS
 
Endogeneity and Entrepreneurship Research
Endogeneity and Entrepreneurship ResearchEndogeneity and Entrepreneurship Research
Endogeneity and Entrepreneurship Research
Brian Anderson
 
Ch9
Ch9Ch9
Mandevillian Intelligence, Paul Smart
Mandevillian Intelligence, Paul SmartMandevillian Intelligence, Paul Smart
Mandevillian Intelligence, Paul Smart
Ulrik Lyngs
 
On the problems of interface: explainability, conceptual spaces, relevance
On the problems of interface: explainability, conceptual spaces, relevanceOn the problems of interface: explainability, conceptual spaces, relevance
On the problems of interface: explainability, conceptual spaces, relevance
Giovanni Sileno
 
Ai4life aiml-xops-sig
Ai4life aiml-xops-sigAi4life aiml-xops-sig
Ai4life aiml-xops-sig
madhucharis
 
Intelligence
IntelligenceIntelligence
Intelligence
lagrada
 
Chapter 7 Lecture Disco 4e
Chapter 7 Lecture Disco 4eChapter 7 Lecture Disco 4e
Chapter 7 Lecture Disco 4e
professorbent
 
First Pages Cotner 907-6
First Pages Cotner 907-6First Pages Cotner 907-6
First Pages Cotner 907-6
bluedoor, LLC.
 
Final Project. Students should prepare a final paper (@ least 8 do
Final Project. Students should prepare a final paper (@ least 8 doFinal Project. Students should prepare a final paper (@ least 8 do
Final Project. Students should prepare a final paper (@ least 8 do
ChereCheek752
 
Week 5 Primary, Secondary data and G
Week 5 Primary, Secondary data and GWeek 5 Primary, Secondary data and G
Week 5 Primary, Secondary data and G
Jamie Davies
 
Week 5 Data Types and Gottesman and Shields 1961
Week 5 Data Types and Gottesman and Shields 1961Week 5 Data Types and Gottesman and Shields 1961
Week 5 Data Types and Gottesman and Shields 1961
Jamie Davies
 
Psych Chapters 1-6 Midterm #1
Psych Chapters 1-6 Midterm #1Psych Chapters 1-6 Midterm #1
Psych Chapters 1-6 Midterm #1
Darrel Adams
 
Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...
Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...
Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...
Innovation and Technology for Development Centre
 
Quals Practice Presentation
Quals Practice PresentationQuals Practice Presentation
Quals Practice Presentation
Vanessa S
 
Building Teams: We Got It All Wrong
Building Teams: We Got It All WrongBuilding Teams: We Got It All Wrong
Building Teams: We Got It All Wrong
Pawel Brodzinski
 
Chapter 11 intelligence
Chapter 11   intelligenceChapter 11   intelligence
Chapter 11 intelligence
swenson_n111
 
Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...
Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...
Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...
jemille6
 
Research methods wccc 9 14-15
Research methods wccc 9 14-15Research methods wccc 9 14-15
Research methods wccc 9 14-15
Ray Brannon
 
Redevelop 2019 - Debugging our biases and intuition in software development
Redevelop 2019 - Debugging our biases and intuition in software developmentRedevelop 2019 - Debugging our biases and intuition in software development
Redevelop 2019 - Debugging our biases and intuition in software development
Dave Hulbert
 

Similar to Detecting emergent intersectional biases: Contextualized word embeddings contain a distribution of human-like biases (20)

From Human Intelligence to Machine Intelligence
From Human Intelligence to Machine IntelligenceFrom Human Intelligence to Machine Intelligence
From Human Intelligence to Machine Intelligence
 
Endogeneity and Entrepreneurship Research
Endogeneity and Entrepreneurship ResearchEndogeneity and Entrepreneurship Research
Endogeneity and Entrepreneurship Research
 
Ch9
Ch9Ch9
Ch9
 
Mandevillian Intelligence, Paul Smart
Mandevillian Intelligence, Paul SmartMandevillian Intelligence, Paul Smart
Mandevillian Intelligence, Paul Smart
 
On the problems of interface: explainability, conceptual spaces, relevance
On the problems of interface: explainability, conceptual spaces, relevanceOn the problems of interface: explainability, conceptual spaces, relevance
On the problems of interface: explainability, conceptual spaces, relevance
 
Ai4life aiml-xops-sig
Ai4life aiml-xops-sigAi4life aiml-xops-sig
Ai4life aiml-xops-sig
 
Intelligence
IntelligenceIntelligence
Intelligence
 
Chapter 7 Lecture Disco 4e
Chapter 7 Lecture Disco 4eChapter 7 Lecture Disco 4e
Chapter 7 Lecture Disco 4e
 
First Pages Cotner 907-6
First Pages Cotner 907-6First Pages Cotner 907-6
First Pages Cotner 907-6
 
Final Project. Students should prepare a final paper (@ least 8 do
Final Project. Students should prepare a final paper (@ least 8 doFinal Project. Students should prepare a final paper (@ least 8 do
Final Project. Students should prepare a final paper (@ least 8 do
 
Week 5 Primary, Secondary data and G
Week 5 Primary, Secondary data and GWeek 5 Primary, Secondary data and G
Week 5 Primary, Secondary data and G
 
Week 5 Data Types and Gottesman and Shields 1961
Week 5 Data Types and Gottesman and Shields 1961Week 5 Data Types and Gottesman and Shields 1961
Week 5 Data Types and Gottesman and Shields 1961
 
Psych Chapters 1-6 Midterm #1
Psych Chapters 1-6 Midterm #1Psych Chapters 1-6 Midterm #1
Psych Chapters 1-6 Midterm #1
 
Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...
Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...
Enfoques de portafolio para la innovación: Navegar por la complejidad de los ...
 
Quals Practice Presentation
Quals Practice PresentationQuals Practice Presentation
Quals Practice Presentation
 
Building Teams: We Got It All Wrong
Building Teams: We Got It All WrongBuilding Teams: We Got It All Wrong
Building Teams: We Got It All Wrong
 
Chapter 11 intelligence
Chapter 11   intelligenceChapter 11   intelligence
Chapter 11 intelligence
 
Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...
Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...
Mathematically Elegant Answers to Research Questions No One is Asking (meta-a...
 
Research methods wccc 9 14-15
Research methods wccc 9 14-15Research methods wccc 9 14-15
Research methods wccc 9 14-15
 
Redevelop 2019 - Debugging our biases and intuition in software development
Redevelop 2019 - Debugging our biases and intuition in software developmentRedevelop 2019 - Debugging our biases and intuition in software development
Redevelop 2019 - Debugging our biases and intuition in software development
 

Recently uploaded

IEEE Aerospace and Electronic Systems Society as a Graduate Student Member
IEEE Aerospace and Electronic Systems Society as a Graduate Student MemberIEEE Aerospace and Electronic Systems Society as a Graduate Student Member
IEEE Aerospace and Electronic Systems Society as a Graduate Student Member
VICTOR MAESTRE RAMIREZ
 
22CYT12-Unit-V-E Waste and its Management.ppt
22CYT12-Unit-V-E Waste and its Management.ppt22CYT12-Unit-V-E Waste and its Management.ppt
22CYT12-Unit-V-E Waste and its Management.ppt
KrishnaveniKrishnara1
 
Properties Railway Sleepers and Test.pptx
Properties Railway Sleepers and Test.pptxProperties Railway Sleepers and Test.pptx
Properties Railway Sleepers and Test.pptx
MDSABBIROJJAMANPAYEL
 
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
IJECEIAES
 
TIME TABLE MANAGEMENT SYSTEM testing.pptx
TIME TABLE MANAGEMENT SYSTEM testing.pptxTIME TABLE MANAGEMENT SYSTEM testing.pptx
TIME TABLE MANAGEMENT SYSTEM testing.pptx
CVCSOfficial
 
4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf
4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf
4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf
Gino153088
 
Computational Engineering IITH Presentation
Computational Engineering IITH PresentationComputational Engineering IITH Presentation
Computational Engineering IITH Presentation
co23btech11018
 
Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...
Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...
Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...
shadow0702a
 
Object Oriented Analysis and Design - OOAD
Object Oriented Analysis and Design - OOADObject Oriented Analysis and Design - OOAD
Object Oriented Analysis and Design - OOAD
PreethaV16
 
Data Driven Maintenance | UReason Webinar
Data Driven Maintenance | UReason WebinarData Driven Maintenance | UReason Webinar
Data Driven Maintenance | UReason Webinar
UReason
 
SCALING OF MOS CIRCUITS m .pptx
SCALING OF MOS CIRCUITS m                 .pptxSCALING OF MOS CIRCUITS m                 .pptx
SCALING OF MOS CIRCUITS m .pptx
harshapolam10
 
An Introduction to the Compiler Designss
An Introduction to the Compiler DesignssAn Introduction to the Compiler Designss
An Introduction to the Compiler Designss
ElakkiaU
 
Gas agency management system project report.pdf
Gas agency management system project report.pdfGas agency management system project report.pdf
Gas agency management system project report.pdf
Kamal Acharya
 
原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样
原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样
原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样
ydzowc
 
Applications of artificial Intelligence in Mechanical Engineering.pdf
Applications of artificial Intelligence in Mechanical Engineering.pdfApplications of artificial Intelligence in Mechanical Engineering.pdf
Applications of artificial Intelligence in Mechanical Engineering.pdf
Atif Razi
 
Engineering Standards Wiring methods.pdf
Engineering Standards Wiring methods.pdfEngineering Standards Wiring methods.pdf
Engineering Standards Wiring methods.pdf
edwin408357
 
一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理
一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理
一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理
upoux
 
Curve Fitting in Numerical Methods Regression
Curve Fitting in Numerical Methods RegressionCurve Fitting in Numerical Methods Regression
Curve Fitting in Numerical Methods Regression
Nada Hikmah
 
Mechanical Engineering on AAI Summer Training Report-003.pdf
Mechanical Engineering on AAI Summer Training Report-003.pdfMechanical Engineering on AAI Summer Training Report-003.pdf
Mechanical Engineering on AAI Summer Training Report-003.pdf
21UME003TUSHARDEB
 
Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...
Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...
Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...
PriyankaKilaniya
 

Recently uploaded (20)

IEEE Aerospace and Electronic Systems Society as a Graduate Student Member
IEEE Aerospace and Electronic Systems Society as a Graduate Student MemberIEEE Aerospace and Electronic Systems Society as a Graduate Student Member
IEEE Aerospace and Electronic Systems Society as a Graduate Student Member
 
22CYT12-Unit-V-E Waste and its Management.ppt
22CYT12-Unit-V-E Waste and its Management.ppt22CYT12-Unit-V-E Waste and its Management.ppt
22CYT12-Unit-V-E Waste and its Management.ppt
 
Properties Railway Sleepers and Test.pptx
Properties Railway Sleepers and Test.pptxProperties Railway Sleepers and Test.pptx
Properties Railway Sleepers and Test.pptx
 
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
 
TIME TABLE MANAGEMENT SYSTEM testing.pptx
TIME TABLE MANAGEMENT SYSTEM testing.pptxTIME TABLE MANAGEMENT SYSTEM testing.pptx
TIME TABLE MANAGEMENT SYSTEM testing.pptx
 
4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf
4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf
4. Mosca vol I -Fisica-Tipler-5ta-Edicion-Vol-1.pdf
 
Computational Engineering IITH Presentation
Computational Engineering IITH PresentationComputational Engineering IITH Presentation
Computational Engineering IITH Presentation
 
Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...
Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...
Use PyCharm for remote debugging of WSL on a Windo cf5c162d672e4e58b4dde5d797...
 
Object Oriented Analysis and Design - OOAD
Object Oriented Analysis and Design - OOADObject Oriented Analysis and Design - OOAD
Object Oriented Analysis and Design - OOAD
 
Data Driven Maintenance | UReason Webinar
Data Driven Maintenance | UReason WebinarData Driven Maintenance | UReason Webinar
Data Driven Maintenance | UReason Webinar
 
SCALING OF MOS CIRCUITS m .pptx
SCALING OF MOS CIRCUITS m                 .pptxSCALING OF MOS CIRCUITS m                 .pptx
SCALING OF MOS CIRCUITS m .pptx
 
An Introduction to the Compiler Designss
An Introduction to the Compiler DesignssAn Introduction to the Compiler Designss
An Introduction to the Compiler Designss
 
Gas agency management system project report.pdf
Gas agency management system project report.pdfGas agency management system project report.pdf
Gas agency management system project report.pdf
 
原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样
原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样
原版制作(Humboldt毕业证书)柏林大学毕业证学位证一模一样
 
Applications of artificial Intelligence in Mechanical Engineering.pdf
Applications of artificial Intelligence in Mechanical Engineering.pdfApplications of artificial Intelligence in Mechanical Engineering.pdf
Applications of artificial Intelligence in Mechanical Engineering.pdf
 
Engineering Standards Wiring methods.pdf
Engineering Standards Wiring methods.pdfEngineering Standards Wiring methods.pdf
Engineering Standards Wiring methods.pdf
 
一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理
一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理
一比一原版(osu毕业证书)美国俄勒冈州立大学毕业证如何办理
 
Curve Fitting in Numerical Methods Regression
Curve Fitting in Numerical Methods RegressionCurve Fitting in Numerical Methods Regression
Curve Fitting in Numerical Methods Regression
 
Mechanical Engineering on AAI Summer Training Report-003.pdf
Mechanical Engineering on AAI Summer Training Report-003.pdfMechanical Engineering on AAI Summer Training Report-003.pdf
Mechanical Engineering on AAI Summer Training Report-003.pdf
 
Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...
Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...
Prediction of Electrical Energy Efficiency Using Information on Consumer's Ac...
 

Detecting emergent intersectional biases: Contextualized word embeddings contain a distribution of human-like biases

  • 1. Detecting Emergent Intersectional Biases:  Contextualized Word Embeddings Contain a Distribution of Human-like Biases Wei Guo Aylin Caliskan August 1, 2021 George Washington University 1
  • 2. Bias in NLP is everywhere! • Bias in NLP perpetuates bias in society • Incomprehensive measurement • Cannot automatically identify bias 2
  • 3. Intersectional bias is concerning! 3
  • 4. Intersectional bias is concerning! • Incomplete measurement of social biases • Unique experiences of discrimination in ML system 3
  • 5. Methods: implicit cognition → natural language → computer vision Implicit Association Test (IAT) • Tests for differential association of two concepts • Easier to categorize stereotype-congruent pairs • Harder to categorize stereotype-incongruent pairs • Effect d = difference in reaction time Weapon IAT (implicit.harvard.edu) 4
  • 6. Methods: implicit cognition → natural language → computer vision Implicit Association Test (IAT) • Tests for differential association of two concepts • Easier to categorize stereotype-congruent pairs • Harder to categorize stereotype-incongruent pairs • Effect d = difference in reaction time Weapon IAT (implicit.harvard.edu) 4
  • 7. Methods: implicit cognition → natural language → computer vision Implicit Association Test (IAT) • Tests for differential association of two concepts • Easier to categorize stereotype-congruent pairs • Harder to categorize stereotype-incongruent pairs • Effect d = difference in reaction time Weapon IAT (implicit.harvard.edu) 4
  • 8. Methods: implicit cognition → natural language → computer vision Implicit Association Test (IAT) • Tests for differential association of two concepts • Easier to categorize stereotype-congruent pairs • Harder to categorize stereotype-incongruent pairs • Effect d = difference in reaction time 4
  • 9. Methods: implicit cognition → static embeddings → contextualized embeddings Word Embedding Association Test Implicit Association Test 5
  • 10. Methods: implicit cognition → static embeddings → contextualized embeddings Word Embedding Association Test 5
  • 11. Methods: implicit cognition → static embeddings → contextualized embeddings man [ feature1 feature2 . . . featured ] father [ feature1 feature2 . . . featured ] . . . woman [ feature1 feature2 . . . featured ] mother [ feature1 feature2 . . . featured ] . . . science [ feature1 feature2 . . . featured ] math [ feature1 feature2 . . . featured ] . . . liberal arts [ feature1 feature2 . . . featured ] music [ feature1 feature2 . . . featured ] . . . 5
  • 12. Methods: implicit cognition → static embeddings → contextualized embeddings man [ feature1 feature2 . . . featured ] father [ feature1 feature2 . . . featured ] . . . woman [ feature1 feature2 . . . featured ] mother [ feature1 feature2 . . . featured ] . . . science [ feature1 feature2 . . . featured ] math [ feature1 feature2 . . . featured ] . . . liberal arts [ feature1 feature2 . . . featured ] music [ feature1 feature2 . . . featured ] . . . Word Embedding Association Test (WEAT) s(w, A, B) = meana∈A cos(w, a)−meanb∈B cos(w, b) s(X, Y, A, B) = ∑ x∈X s(x, A, B) − ∑ y∈Y s(y, A, B) 5
  • 13. Methods: implicit cognition → static embeddings → contextualized embeddings Implicit Association Test Word Embedding Factual Association Test (WEFAT) s(w, A, B) = meana∈As(⃗ w,⃗ a) − meanb∈Bs(⃗ w,⃗ b) stdx∈A∪Bs(⃗ w, x) 6
  • 14. Intersectional Bias Detection (IBD) s(w, A, B) = meana∈As(⃗ w,⃗ a) − meanb∈Bs(⃗ w,⃗ b) stdx∈A∪Bs(⃗ w, x) 7
  • 15. Intersectional Bias Detection (IBD) s(w, A, B) = meana∈As(⃗ w,⃗ a) − meanb∈Bs(⃗ w,⃗ b) stdx∈A∪Bs(⃗ w, x) Detecting intersectional biases associated with members of multiple minority groups. 7
  • 16. Emergent Intersectional Bias Detection (EIBD) 8
  • 17. Emergent Intersectional Bias Detection (EIBD) Intersectional biases - Attributes highly associated with single social category = Remaining set is the emergent intersectional biases Detecting unique emergent intersectional biases that do not overlap with the biases of their constituent minority identities.  8
  • 18. Evaluation of IBD Detection accuracy > 80% accuracy, where random chance < 15% Validation set for intersectional biases from Ghavami Peplau, 2013 9
  • 19. Methods: implicit cognition → static embeddings → contextualized embeddings Extract the sentence containing the words X, Y, A, B Contextualized Embedding Assoiciation Test (CEAT) 10
  • 20. Methods: implicit cognition → static embeddings → contextualized embeddings Generates the contextualized embeddings 10
  • 21. Methods: implicit cognition → static embeddings → contextualized embeddings Calculate the effect size of bias based on WEAT 10
  • 22. Contextualized Embedding Association Test (CEAT) Generates the distribution of effect magnitudes of biases Calculate Combined Effect Size CES(X, Y, A, B) = ∑N i=1 viESi ∑N i=1 vi 10
  • 23. Evaluation of CEAT Contextualized embeddings from Corpus of • Widely shared biases • Flowers/insects • Musical instru- ments/weapons • Social group biases • Gender • Race • Intersectionality • ... 11
  • 24. Evaluation of CEAT • Intersectional biases have high magnitude. • Biased: ELMo > BERT > GPT > GPT-2 • The overall magnitude of bias negatively correlates with the level of contextualization in the language model. 11