A framework for the computational linguistic analysis of - - PowerPoint PPT Presentation

▶
a framework for the computational linguistic analysis of
SMART_READER_LITE
LIVE PREVIEW

A framework for the computational linguistic analysis of - - PowerPoint PPT Presentation

Aug 02, 2023 230 likes •641 views

A framework for the computational linguistic analysis of dehumanization Julia Mendelsohn Dan Jurafsky Yulia Tsvetkov juliamendelsohn.github.io stanford.edu/~jurafsky cs.cmu.edu/~ytsvetko @jmendelsohn2 @jurafsky ytsvetko@cs.cmu.edu 1

slide-1
SLIDE 1

IC2S2 | July 19, 2020

A framework for the computational linguistic analysis of dehumanization

1

Yulia Tsvetkov

cs.cmu.edu/~ytsvetko ytsvetko@cs.cmu.edu

Julia Mendelsohn

juliamendelsohn.github.io @jmendelsohn2 juliame@umich.edu

Dan Jurafsky

stanford.edu/~jurafsky @jurafsky jurafsky@stanford.edu

slide-2
SLIDE 2

IC2S2 | July 19, 2020

Dehumanization

The act of perceiving or treating people as less than human [Haslam & Stratemeyer, 2016]

2

(https://cdn.psychologytoday.com)

Leads to extreme intergroup bias, hate speech, violence

slide-3
SLIDE 3

IC2S2 | July 19, 2020

  • We identify linguistic analogs for several dimensions
  • f dehumanization and propose computational

techniques to measure these linguistic correlates.

  • Case Study: changing representations of LGBTQ

groups in the New York Times over three decades.

  • Through this lens, we investigate differences in social

meaning between seemingly similar group labels.

3

This talk

slide-4
SLIDE 4

IC2S2 | July 19, 2020

Components of dehumanization

  • 1. Negative evaluations of the target group
  • 2. Moral disgust
  • 3. Associations with non-humans (especially vermin)
  • 4. Denial of agency
  • 5. Psychological distance
  • 6. Essentialism
  • 7. Denial of subjectivity

4

slide-5
SLIDE 5

IC2S2 | July 19, 2020

Components of dehumanization

  • 1. Negative evaluations of the target group
  • 2. Moral disgust
  • 3. Associations with non-humans (especially vermin)

We operationalize these three components by identifying and measuring lexical semantic analogs.

5

slide-6
SLIDE 6

IC2S2 | July 19, 2020

Components of dehumanization

  • 1. Negative evaluations of the target group

6

Attribution of negative characteristics to target group categorizes groups that are “excluded from the realm of acceptable norms and values” [Bar-Tal, 1990]

slide-7
SLIDE 7

IC2S2 | July 19, 2020

Components of dehumanization

  • 1. Extremely negative evaluations
  • 2. Moral Disgust

7

Disgust → perception of target group’s negative social value [Sherman & Haidt, 2011] Moral disgust “facilitates moral exclusion of

  • ut-groups” [Buckels & Trapnell, 2013]
slide-8
SLIDE 8

IC2S2 | July 19, 2020

Components of dehumanization

  • 1. Extremely negative evaluations
  • 2. Denial of agency
  • 3. Associations with non-humans (especially vermin)

8

Vermin metaphor conceptualizes the target group as “engaged in threatening behavior, but devoid of thought or emotional desire” [Tipler & Ruscher, 2014]

slide-9
SLIDE 9

IC2S2 | July 19, 2020

Quantifying negative evaluations (1)

Valence: aspect of meaning ranging from negative emotion (unpleasant) to positive (pleasant) NRC VAD lexicon: valence scores from 0 to 1 for 20k English words

9

Word Score love 1.000 happy 1.000 happily 1.000 toxic 0.008 nightmare 0.005 shit 0.000

Obtaining Reliable Human Ratings of Valence, Arousal, and Dominance for 20,000 English Words. Mohammad,S. (2018). ACL.

slide-10
SLIDE 10

IC2S2 | July 19, 2020

Quantifying negative evaluations (2)

The cosine similarity between words in vector space models reflects similarity in meaning We estimate a group label’s valence by training word vectors, measuring average valence over the label’s nearest K neighbors

10

Hamilton, WL, et al. (2016). Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change. ACL.

slide-11
SLIDE 11

IC2S2 | July 19, 2020

Quantifying moral disgust

Create vector representation for Moral Disgust Concept Weighted average of word vectors from Moral Foundations Dict (46 words/stems) Cosine similarity between Moral Disgust Concept and group label

11

disgust* sin filth* gross repuls* pervert profan*

  • bscen*

Graham, J., Haidt, J., & Nosek, B. A. (2009). Liberals and conservatives rely on different sets of moral foundations..

Moral Disgust Concept Group Label

slide-12
SLIDE 12

IC2S2 | July 19, 2020

Quantifying vermin metaphors

Create vector representation for Vermin Concept Weighted average of verminy word vectors Cosine similarity between Vermin Concept and group label

12

vermin rodent(s) rat(s) cockroach(es) mice termite(s) fleas bedbug(s)

Vermin Concept Group Label

slide-13
SLIDE 13

IC2S2 | July 19, 2020

Methods Summary

13

Dehumanization Element Operationalization Negative evaluation of target group Paragraph-level valence analysis Connotation frames of perspective Word embedding neighbor valence Denial of agency Connotation frames of agency Word embedding neighbor agency Moral disgust Vector similarity to disgust Vermin metaphor Vector similarity to vermin

slide-14
SLIDE 14

IC2S2 | July 19, 2020

LGBTQ representation in the New York Times

Americans have become more supportive of LGBTQ rights, but LGBTQ people still face significant discrimination Homosexual: outdated label with clinical and sexual associations

14

slide-15
SLIDE 15

IC2S2 | July 19, 2020

Data

Word embeddings trained per year on full NYT 1986-2015

15

slide-16
SLIDE 16

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

16

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-17
SLIDE 17

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

17

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-18
SLIDE 18

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

18

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-19
SLIDE 19

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

19

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-20
SLIDE 20

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

20

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-21
SLIDE 21

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

21

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-22
SLIDE 22

IC2S2 | July 19, 2020

Word embedding top nearest neighbors

22

1986 2015 gay homosexual gay homosexual homophobia premarital interracial premarital women sexual couples bestiality feminist promiscuity marriage pedophilia suffrage polygamy closeted adultery sexism anal equality infanticide a.c.l.u. intercourse abortion abhorrent amen consenting unmarried feticide queer consensual

  • penly

fornication

slide-23
SLIDE 23

IC2S2 | July 19, 2020

Results: negative evaluations

  • Evaluations of LGBTQ

people have improved

  • ver time
  • Homosexual associated

with more negative words than gay

23

  • Homosexual’s neighboring words become more

negative, suggesting that this term is used in more negative (and potentially dehumanizing) contexts

slide-24
SLIDE 24

IC2S2 | July 19, 2020

Results: moral disgust & vermin metaphor

  • LGBTQ terms have become less associated with moral

disgust and vermin over time

  • Homosexual is more associated with moral disgust

and vermin than gay, especially after 2000

24

slide-25
SLIDE 25

IC2S2 | July 19, 2020

Summary

Our framework involves:

  • 1. Identifying aspects of dehumanization from literature
  • 2. Measuring lexical semantic correlates with

computational methods

  • 3. Qualitative & quantitative evaluation (in paper)

Our study of LGBTQ representation in the NYT revealed:

  • Increasingly humanizing descriptions of LGBTQ people
  • Homosexual emerged as an index of more

dehumanizing attitudes than other terms (esp. gay)

25

slide-26
SLIDE 26

IC2S2 | July 19, 2020

Interdisciplinary Contributions

Framework for large-scale study of dehumanization Linguistics: language variation and change in discourses surrounding marginalized social groups Psych: complement small-scale dehumanization studies CS: Detection of media bias and abusive language

26

slide-27
SLIDE 27

IC2S2 | July 19, 2020

Thank you!

A preprint of our paper is available here.

27

Julia Mendelsohn

juliamendelsohn.github.io @jmendelsohn2 juliame@umich.edu

Dan Jurafsky

stanford.edu/~jurafsky @jurafsky jurafsky@stanford.edu

Yulia Tsvetkov

cs.cmu.edu/~ytsvetko ytsvetko@cs.cmu.edu

slide-28
SLIDE 28

IC2S2 | July 19, 2020

References

  • Bar-Tal, D. (1990). Causes and consequences of delegitimization: Models of conflict and ethnocentrism. Journal of Social issues 46, 65–81
  • Buckels, E. E. and Trapnell, P. D. (2013). Disgust facilitates outgroup dehumanization. Group Processes & Intergroup Relations 16, 771–780
  • Gallup (2019). Gay and lesbian rights. http://news.gallup.com/poll/1651/gay-lesbian-rights.aspx
  • Graham, J., Haidt, J., and Nosek, B. A. (2009). Liberals and conservatives rely on different sets of moral foundations. Journal of personality and social

psychology 96-1029

  • Hamilton, W. L., Leskovec, J., & Jurafsky, D. (2016). Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change. In Proceedings of the 54th Annual

Meeting of the Association for Computational Linguistics (ACL)

  • Haslam, N. and Stratemeyer, M. (2016). Recent research on dehumanization. Current Opinion in Psychology 11, 25–29
  • Mohammad, S. M. (2018). Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 english words. In Proceedings of The 56th Annual

Meeting of the Association for Computational Linguistics (ACL)

  • Opotow, S. (1990). Moral exclusion and injustice: An introduction. Journal of social issues 46, 1–20
  • Rashkin, H., Singh, S., and Choi, Y. (2016). Connotation frames: A data-driven investigation. In Proceedings of the 54th Annual Meeting of the Association for

Computational Linguistics

  • Sap, M., Prasettio, M. C., Holtzman, A., Rashkin, H., and Choi, Y. (2017). Connotation frames of power and agency in modern films. In Proceedings of the 2017

Conference on Empirical Methods in Natural Language Processing. 2329–2334

  • Sherman, G. D. and Haidt, J. (2011). Cuteness and disgust: the humanizing and dehumanizing effects of emotion. Emotion Review 3, 245–251
  • Tipler, C. and Ruscher, J. B. (2014). Agency’s role in dehumanization: Non-human metaphors of out-groups. Social and Personality Psychology Compass 8,

214–228 28

slide-29
SLIDE 29

IC2S2 | July 19, 2020

Additional slides

29

slide-30
SLIDE 30

IC2S2 | July 19, 2020

Bias in human-annotated VAD lexicon

We filtered LGBTQ labels before calculating valence

30

LGBTQ term Valence Other term Valence transsexual 0.264 woman 0.865 homosexual 0.333 human 0.767 lesbian 0.385 man 0.688 gay 0.388 person 0.646 bisexual 0.438 heterosexual 0.561

slide-31
SLIDE 31

IC2S2 | July 19, 2020

Quantifying negative evaluations (1)

Valence: evaluation from negative (unpleasant) to positive (pleasant) NRC VAD lexicon: valence scores from 0 to 1 for 20k English words Calculate average valence score

  • ver all words in the text

31

Word Score love 1.000 happy 1.000 happily 1.000 toxic 0.008 nightmare 0.005 shit 0.000

Obtaining Reliable Human Ratings of Valence, Arousal, and Dominance for 20,000 English Words. Mohammad,S. (2018). ACL.

slide-32
SLIDE 32

IC2S2 | July 19, 2020

Quantifying negative evaluations (2)

We want to measure valence directed towards target group Connotation Frames Lexicon: 900 verbs, writer’s perspective towards subj and obj Extracted SVO tuples for head verbs where group label was in subj or obj NP

32

X Y Writer

P(w → X) = -- P(w → Y) = + P(X → Y) = --

Rashkin, H., Singh, S., & Choi, Y. (2016). Connotation Frames: A Data-Driven Investigation. ACL.

X violates Y

slide-33
SLIDE 33

IC2S2 | July 19, 2020

Components of dehumanization

  • 1. Extremely negative evaluations
  • 2. Denial of agency
  • 3. Moral disgust
  • 4. Denial of agency

Agency: The ability to: (1) experience emotion & feel pain (affective mental states) (2) act & produce effect on environment (behavioral potential) (3) think & hold beliefs (cognitive mental states)

[Tipler & Ruscher, 2014]

33

slide-34
SLIDE 34

IC2S2 | July 19, 2020

Quantifying denial of agency

The man beckons Irene forward He obeys, eyes bulging

34

Sap, M. et al. (2017). Connotation frames of power and agency in modern films. EMNLP.

+ agency

  • agency

Agency Connotation Frames: 2k verbs labeled for agency High agency: high control, active decision-makers Low agency: more passive Fraction of high-agency subjects in SV pairs containing group label

slide-35
SLIDE 35

IC2S2 | July 19, 2020

Quantifying denial of agency (2)

NRC VAD lexicon: dominance scores from 0 to 1 for 20k words Calculate dominance score over nearest K word2vec neighbors Limitation: power != agency

35

Word Score powerful 0.991 leadership 0.983 success 0.981 empty 0.081 frail 0.069 weak 0.045

Obtaining Reliable Human Ratings of Valence, Arousal, and Dominance for 20,000 English Words. Mohammad,S. (2018). ACL.

slide-36
SLIDE 36

IC2S2 | July 19, 2020

Tradeoffs: negative evaluation methods

36

Paragraph Connotation frames Vector neighbors interpretable interpretable less interpretable broader context limited scope broader context not directed directed directed topical effects syntax is hard major events Disentangling perspectives within text

slide-37
SLIDE 37

IC2S2 | July 19, 2020

37

slide-38
SLIDE 38

IC2S2 | July 19, 2020

Future directions

  • More sophisticated methods (contextual embeddings)
  • Measure other dimensions of dehumanization and

non-lexical semantic cues

○ Denial of subjectivity (quote attribution) ○ Psychological distance (definite plurals) ○ Essentialism (noun v. adjective forms)

  • Other LGBTQ terms, groups, data sources, languages

38

slide-39
SLIDE 39

IC2S2 | July 19, 2020

Ethical concerns

  • Biases in lexicons and methods
  • Vectors are dehumanizing
  • Case Study: Aggregated LGBTQ representations

suppress diversity of identities within this umbrella

  • Emphasis on gay and homosexual and erasure of

marginalized people within LGBTQ communities

39