Hate Speech Classifiers Learn Normative Social Stereotypes
Aida Mostafazadeh Davani, Mohammad Atari, Brendan Kennedy, Morteza Dehghani
University of Southern California, USA
{mostafaz,atari,btkenned,mdehghan}@usc.edu
Abstract
Social stereotypes negatively impact individ-
uals’ judgments about different groups and
may have a critical role in understanding lan-
guage directed toward marginalized groups.
Here, we assess the role of social stereotypes
in the automated detection of hate speech
in the English language by examining the
impact of social stereotypes on annotation
behaviors, annotated datasets, and hate speech
classifiers. Specifically, we first investigate the
impact of novice annotators’ stereotypes on
their hate-speech-annotation behavior. Then,
we examine the effect of normative stereo-
types in language on the aggregated annota-
tors’ judgments in a large annotated corpus.
Finally, we demonstrate how normative stereo-
types embedded in language resources are
associated with systematic prediction errors
in a hate-speech classifier. The results dem-
onstrate that hate-speech classifiers reflect so-
cial stereotypes against marginalized groups,
which can perpetuate social inequalities when
propagated at scale. This framework, combin-
ing social-psychological and computational-
linguistic methods, provides insights into
sources of bias in hate-speech moderation,
informing ongoing debates regarding machine
learning fairness.
Introduction
Artificial Intelligence (AI) technologies are prone
to acquiring cultural, social, and institutional bi-
ases from the real-world data on which they are
trained (McCradden et al., 2020; Mehrabi et al.,
2021; Obermeyer et al., 2019). AI models trained
on biased datasets both reflect and amplify those
biases (Crawford, 2017). For example, the domi-
nant practice in modern Natural Language Pro-
cessing (NLP)—which is to train AI systems
on large corpora of human-generated text data—
leads to representational biases, such as preferring
European American names over African Amer-
ican names (Caliskan et al., 2017), associating
words with more negative sentiment with phrases
300
referencing persons with disabilities (Hutchinson
et al., 2020), making ethnic stereotypes by asso-
ciating Hispanics with housekeepers and Asians
with professors (Garg et al., 2018), and assign-
ing men to computer programming and women
to homemaking (Bolukbasi et al., 2016).
Moreover, NLP models are particularly suscep-
tible to amplifying biases when their task involves
evaluating language generated by or describing a
social group (Blodgett and O’Connor, 2017). For
example, previous research has shown that tox-
icity detection models associate documents con-
taining features of African American English with
higher offensiveness than text without those fea-
tures (Sap et al., 2019; Davidson et al., 2019).
Similarly, Dixon et al. (2018) demonstrate that
models trained on social media posts are prone
to erroneously classifying ‘‘I am gay’’ as hate
speech. Therefore, using such models for moder-
ating social-media platforms can yield dispropor-
tionate removal of social-media posts generated
by or mentioning marginalized groups (Davidson
et al., 2019). This unfair assessment negatively
impacts marginalized groups’ representation in
online platforms, which leads to disparate impacts
on historically excluded groups (Feldman et al.,
2015).
Mitigating biases in hate speech detection, nec-
essary for viable automated content moderation
(Davidson et al., 2017; Mozafari et al., 2020), has
recently gained momentum (Davidson et al., 2019;
Dixon et al., 2018; Sap et al., 2019; Kennedy et al.,
2020; Prabhakaran et al., 2019). Most current su-
pervised algorithms for hate speech detection rely
on data resources that potentially reflect real-
world biases: (1) text representation, which maps
textual data to their numeric representations in a
semantic space; and (2) human annotations, which
represent subjective judgments about
the hate
speech content of the text, constituting the train-
ing dataset. Both (1) and (2) can introduce biases
into the final model. First, a classifier may be-
come biased due to how the mapping of language
Transactions of the Association for Computational Linguistics, vol. 11, pp. 300–319, 2023. https://doi.org/10.1162/tacl a 00550
Action Editor: Alice Oh. Submission batch: 4/2022; Revision batch: 11/2022; Published 3/2023.
c(cid:2) 2023 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license.
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
to numeric representations is affected by stereo-
typical co-occurrences in the training data of the
language model. For example, a semantic asso-
ciation between phrases referencing persons with
disabilities and words with more negative senti-
ment in the language model can impact a classi-
fier’s evaluation of a sentence about disability
(Hutchinson et al., 2020). Second, individual-level
biases of annotators can impact the classifier in
stereotypical directions. For example, a piece of
rhetoric about disability can be analyzed and
labeled differently depending upon annotators’
social biases.
Although previous research has documented
stereotypes in text representations (Garg et al.,
2018; Bolukbasi et al., 2016; Manzini et al., 2019;
Swinger et al., 2019; Charlesworth et al., 2021),
the impact of annotators’ biases on training data
and models remains largely unknown. Filling this
gap in our understanding of the effect of human
annotation on biased NLP models is the focus of
this work. As argued by Blodgett et al. (2020)
and Kiritchenko et al. (2021), a comprehensive
evaluation of human-like biases in hate speech
classification needs to be grounded in social psy-
chological theories of prejudice and stereotypes,
in addition to how they are manifested in lan-
guage. In this paper, we rely on the Stereotype
Content Model (SCM; Fiske et al., 2002) which
suggests that social perceptions and stereotyp-
ing form along two dimensions, namely, warmth
(e.g., trustworthiness, friendliness) and compe-
tence (e.g., capability, assertiveness). The SCM’s
main tenet is that perceived warmth and compe-
tence underlie group stereotypes. Hence, different
social groups can be positioned in different loca-
tions in this two-dimensional space, since much
of the variance in stereotypes of groups is ac-
counted for by these basic social psychological
dimensions.
In three studies presented in this paper, we study
the pipeline for training a hate speech classi-
fier, consisting of collecting annotations, aggre-
gating annotations for creating the training dataset,
and training the model. We investigate the effects
of social stereotypes on each step, namely, (1)
the relationship between social stereotypes and
hate speech annotation behaviors, (2) the relation-
ship between social stereotypes and aggregated
annotations of trained, expert annotators in cu-
rated datasets, and (3) social stereotypes as they
manifest in the biased predictions of hate speech
classifiers. Our work demonstrates that differ-
ent stereotypes along warmth and competence
differentially affect
individual annotators, cu-
rated datasets, and trained language classifiers.
Therefore, understanding the specific social bi-
ases targeting different marginalized groups is
essential for mitigating human-like biases of AI
models.
1 Study 1: Text Annotation
Here, we investigate the effect of individuals’
social stereotypes on their hate speech annotations.
Specifically, we aim to determine whether novice
annotators’ stereotypes (perceived warmth and/or
competence) of a mentioned social group lead
to higher rate of labeling text as hate speech and
higher rates of disagreement with other annotators.
We conduct a study on a nationally stratified
sample (in terms of age, ethnicity, gender, and
political orientation) of US adults. First, we ask
participants to rate eight US-relevant social groups
on different stereotypical traits (e.g., friendliness).
Then, participants are presented with social media
posts mentioning the social groups and are asked
to label the content of each post based on whether
it attacks the dignity of that group. We expect
the perceived warmth and/or competence of the
social groups to be associated with participants’
annotation behaviors, namely, their rate of labeling
text as hate speech and disagreeing with other
annotators.
Participants To achieve a diverse set of an-
notations, we recruited a relatively large (N =
1,228) set of participants in a US sample stratified
across participants’ gender, age, ethnicity, and po-
litical ideology through Qualtrics Panels.1 After
filtering participants based on quality-check items
(described below), our final sample included 857
American adults (381 male, 476 female) rang-
ing in age from 18 to 70 (M = 46.7, SD =
16.4) years, about half Democrats (50.4%) and
half Republicans (49.6%), with diverse reported
race/ethnicity (67.8% White or European Amer-
ican, 17.5% Black or African American, 17.7%
Hispanic or Latino/Latinx, 9.6% Asian or Asian
American).
1https://www.census.gov/quickfacts/fact
/table/US/PST045221.
301
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Stimuli To compile a set of stimuli items for this
study, we selected posts from the Gab Hate Cor-
pus (GHC; Kennedy et al., 2022), which includes
27,665 social-media posts collected from the cor-
pus of Gab.com (Gaffney, 2018), each annotated
for their hate speech content by at least three ex-
pert annotators. We collected all posts with high
disagreement among the GHC’s (original) anno-
tators (based on Equation 1 for quantifying item
disagreement) which mention at least one social
group. We searched for posts mentioning one of
the eight most frequently targeted social groups
in the GHC: (1) women; (2) immigrants; (3) Mus-
lims; (4) Jews; (5) communists; (6) liberals; (7)
African Americans; and (8) homosexual individ-
uals. We selected seven posts per group, result-
ing in a set of 56 items in total.
Explicit Stereotype Measure We assessed par-
ticipants’ warmth and competence stereotypes of
the 8 US social groups in our study based on
their perceived traits for a typical member of each
group. To this end, we followed social psycholog-
ical approaches for collecting these self-reported,
explicit stereotypes (Cuddy et al., 2008) and
asked participants to rate a typical member of
each social group (e.g., Muslims) based on their
‘‘friendliness’’, ‘‘helpfulness,’’ ‘‘peacefulness,’’
and ‘‘intelligence.’’ Following previous studies
of perceived stereotypes (Huesmann et al., 2012;
Cuddy et al., 2007), participants were asked to
rate these traits from low (e.g., ‘‘unfriendly’’) to
high (e.g., ‘‘friendly’’) using an 8-point semantic
differential scale. We considered the average of
the first three traits as the indicator of perceived
warmth2 and the fourth item as the perceived
competence.
While explicit assessments are generally corre-
lated with implicit measures of attitude, in the case
of self-reporting social stereotypes, participants’
explicit answers can be less significantly corre-
lated with their implicit biases, potentially due
to motivational and cognitive factors (Hofmann
et al., 2005). Therefore, it should be noted that
this study relies on an explicit assessment of so-
cial stereotypes, and the results do not directly
explain the effects of implicit biases on annotat-
ing hate speech.
2Cronbach’s α’s ranged between .90 [women] and .95
[Muslims].
Hate Speech Annotation Task We asked par-
ticipants to annotate the 56 items based on a short
definition of hate speech (Kennedy et al., 2022):
‘‘Language that intends to attack the dignity of
a group of people, either through an incitement
to violence, encouragement of the incitement to
violence, or the incitement to hatred.’’
Participants could proceed with the study only
after they acknowledged understanding the pro-
vided definition of hate speech. We then tested
their understanding of the definition by plac-
ing three synthetic ‘‘quality-check’’ items among
survey items, two of which included clear and
explicit hateful language directly matching our
definition and one item that was simply infor-
mational (see Supplementary Materials). Overall,
371 out of the original 1,228 participants failed
to satisfy these conditions and their input was
removed from the data.3
Disagreement Throughout this paper, we as-
sess annotation disagreement in different levels:
• Item disagreement, d(i): Motivated by Fleiss
(1971), for each item i, item disagreement
d(i) is the number of annotator pairs that
disagree on the item’s label, divided by the
number of all possible annotator pairs.4
d(i) =
× n(i)
n(i)
1
0
(cid:3)
(cid:2)
1 +n(i)
n(i)
2
0
(1)
and n(i)
0
Here, n(i)
show the number of
1
hate and non-hate labels assigned to i,
respectively.
• Participant item-level disagreement, d(p,i):
For each participant p and each item i, we
define d(p,i) as the ratio of participants with
whom p agreed, to the size of the set of par-
ticipants who annotated the same item (P ).
d(p,i) =
|{p(cid:3)|p(cid:3) ∈ P − {p}, yp,i = yp(cid:3),i}|
|P |
(2)
Here, yp,i is the label that p assigned to i.
3The replication of our analyses with all participants
yielded similar results, reported in Supplementary Materials.
4We found this measure more suitable than a simple
percentage, as Fleiss captures the total number of annotators
as well as the disagreeing pairs.
302
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
• Group-level disagreement, d(p,S): For a spe-
cific set of items S and an annotator p, d(p,S)
captures how much p disagrees with others
over items in S. We calculate d(p,S) by
averaging d(p,i)s for all items i ∈ S
d(p,S) =
1
|S|
(cid:4)
i∈S
d(p,i)
(3)
Annotators’ Tendency To explore participants’
annotation behaviors relative to other participants,
we rely on the Rasch model (Rasch, 1993). The
Rasch model is a psychometric method that mod-
els participants’ responses—here, annotations—
to items by calculating two sets of parameters,
namely, the ability of each participant and the
difficulty of each item. Similar approaches, based
on Item Response Theory (IRT), have recently
been applied in evaluating NLP models (Lalor
et al., 2016) and for modeling the relative per-
formance of annotators (Hovy et al., 2013).
While, compared to Rasch models, IRT mod-
els can include more item-level parameters, our
choice of Rasch models is based on their ro-
bust estimations for annotators’ ability scores.
Specifically, Rasch models calculate the ability
score solely based on individuals’ performances
and independent from the sample set. In contrast,
in IRT-based approaches, individual annotators’
scores depend on the complete set of annotators
(Stemler and Naples, 2021). To provide an estima-
tion of these two sets of parameters (annotators’
ability and items’ difficulty), the Rasch model
iteratively fine-tunes parameters’ values to ulti-
mately fit the best probability model to partici-
pants’ responses to items. Here, we apply a Rasch
model to each set of items mentioning a specific
social group.
It should be noted that Rasch models con-
sider each response as either correct or incorrect
and estimate participants’ ability and items’ diffi-
culty based on the underlying logic that subjects
have a higher probability of correctly answering
easier items. However, we assume no ‘‘ground
truth’’ for the labels, therefore ‘‘1’’s and ‘‘0’’s
represent annotators ‘‘hate’’ and ‘‘not hate’’
answers. Therefore, items’ difficulty (which orig-
inally represents the probability of ‘‘0’’ labels)
can be interpreted as non-hatefulness (probabil-
ity of ‘‘non-hate’’ labels). Respectively, partici-
pants’ ability (probability of getting a ‘‘1’’ for a
difficult item), can be interpreted as their ten-
Figure 1: The overview of Study 1. Novice annotators
are asked to label hate speech content of each post.
Then, their annotation behaviors, per social group to-
ken, are taken to be the number of posts they labeled as
hate speech, their disagreement with other annotators
and their tendency to identify hate speech.
dency towards labeling text as hate (labeling
non-hateful items as hateful). Throughout this
study we use tendency to refer to the ability
parameter.
Analysis We estimate associations between par-
ticipants’ social stereotypes about each social
group with their annotation behaviors evaluated
on items mentioning that social group. Namely,
the dependent variables are (1) the number of hate
labels, (2) the tendency (via the Rasch model) to
detect hate speech relative to others, and (3) the
ratio of disagreement with other participants—as
quantified by group-level disagreement. To ana-
lyze annotation behaviors concerning each social
group, we considered each pair of participant
(N = 857) and social group (ngroup = 8) as
an observation (ntotal = 6,856). Each observa-
tion includes the social group’s perceived warmth
and competence based on the participant’s an-
swer to the explicit stereotype measure, as well
as their annotation behaviors on items that men-
tion that social group. Since each observation
is nested in and affected by annotator-level and
social-group level variable, we fit cross-classified
multi-level models to analyze the association
of annotation behaviors with social stereotypes.
Figure 1 illustrates our methodology in con-
ducting Study 1. All analyses were performed in
303
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
R (3.6.1), and the eRm (1.0.1) package was used
for the Rasch model.
Results We first investigated the relation be-
tween participants’ social stereotypes about each
social group and the number of hate speech la-
bels they assigned to items mentioning that group.
The result of a cross-classified multi-level Poisson
model, with the number of hate speech labels as
the dependent variable and participants’ percep-
tion of warmth and competence as independent
variables, shows that a higher number of items
are categorized as hate speech when participants
perceive that social group as high on competence
(β = 0.03, SE = 0.006, p < .001). In other words,
a one point increase in a participant’s rating of a
social group’s competence (on the scale of 1 to 8)
is associated with a 3.0% increase in the number
of hate labels they assigned to items mentioning
that social group. Perceived warmth scores were
not significantly associated with the number of
hate labels (β = 0.01, p = .128).
We then compared annotators’ relative ten-
dency to assign hate speech labels to items
mentioning each social group, calculated by the
Rasch models. We conducted a cross-classified
multi-level linear model to predict participants’
tendency as the dependent variable, and each
social group’s warmth and competence stereo-
types as independent variables. The result shows
that participants demonstrate higher tendency (to
assign hate speech labels) on items that men-
tion a social group they perceive as highly
competent (β = 0.07, SE = 0.013, p < .001).
However, perceived warmth scores were not sig-
nificantly associated with participants’ tendency
scores (β = 0.02, SE = 0.014, p = 0.080).
Finally, we analyzed participants’ group-level
disagreement for items that mention each social
group. We use a logistic regression model to pre-
dict disagreement ratio, which is a value between 0
and 1. The results of a cross-classified multi-level
logistic regression, with group-level disagreement
ratio as the dependent variable and warmth and
competence stereotypes as independent variables,
show that participants disagreed more on items
that mention a social group which they perceive
as low on competence (β = −0.29, SE = 0.001,
p < .001). In other words, a one point decrease
in a participant’s rating of a social group’s com-
petence (on the scale of 1 to 8) is associated with
a 25.2% increase in their odds of disagreement
Figure 2: The relationship between the stereotypical
competence of social groups and (1) the number of
hate labels annotators detected, (2) their tendency to
detect hate speech, and (3) their ratio of disagree-
ment with other participants (top to bottom).
on items mentioning that social group. Perceived
warmth scores were not significantly associated
with the odds of disagreement (β = 0.05, SE =
0.050, p = .322).
In summary, as represented in Figure 2, the
results of Study 1 demonstrate that when novice
annotators perceive a mentioned social group as
high on competence they (1) assign more hate
speech labels, (2) show higher tendency for iden-
tifying hate speech, and (3) disagree less with
other annotators. These associations collectively
denote that when annotators stereotypically per-
ceive a social group as highly competent, they
tend to become more sensitive or alert about hate
speech directed toward that group. These results
support the idea that hate speech annotation is af-
fected by annotators’ stereotypes (specifically the
perceived competence) of target social groups.
2 Study 2: Ground-Truth Generation
The high levels of inter-annotator disagreements
in hate speech annotation (Ross et al., 2017) can
304
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
the inter-annotator disagreement and the majority
vote for each of the posts and considered them as
dependent variables in our analyses.
Quantifying Social Stereotypes To quantify
social stereotypes about each social group from
our list of social group tokens (Dixon et al.,
2018), we calculated semantic similarity of that
social group term with lexicons (dictionaries) of
competence and warmth (Pietraszkiewicz et al.,
2019). The competence and warmth dictionaries
consist of 192 and 184 tokens, respectively, and
have been shown to measure linguistic markers
of competence and warmth reliably in different
contexts.
We calculated the similarity of each social
group token with the entirety of words in dictio-
naries of warmth and competence in a latent vec-
tor space based on previous approaches (Caliskan
et al., 2017; Garg et al., 2018). Specifically, for
each social group token, s and each word w in
the dictionaries of warmth (Dw) or competence
(Dc) we first obtain their numeric representation
(R(s) ∈ Rt and R(w) ∈ Rt, respectively) from
pre-trained English word embeddings (GloVe;
Pennington et al., 2014). The representation func-
tion, R(), maps each word to a t-dimensional
vector, trained based on the word co-occurrences
in a corpus of English Wikipedia articles. Then,
the warmth and competence scores for each so-
cial group token were calculated by averaging
the cosine similarity of the numeric represen-
tation of the social group token and the numeric
representation of the words of the two dictionaries.
Results We examined the effects of the quanti-
fied social stereotypes on hate speech annotations
captured in the dataset. Specifically, we compared
post-level annotation disagreements with the men-
tioned social group’s warmth and competence.
For example, based on this method, ‘‘man’’ is the
most semantically similar social group token to
the dictionary of competence (Cman = 0.22),
while ‘‘elder’’ is the social group token with the
closest semantic representation to the dictionary
of warmth (Welder = 0.19). Of note, we inves-
tigated the effect of these stereotypes on hate
speech annotation of social media posts that men-
tion at least one social group token (Nposts =
5535). Since some posts mention more than one
social group token, we considered each men-
tioned social group token as an observation
Figure 3: The overview of Study 2. We investigate
a hate speech dataset and evaluate the inter-annotator
disagreement and majority label for each document in
relation to stereotypes about mentioned social groups.
be attributed to numerous factors, including an-
notators’ varying perception of the hateful lan-
guage, or ambiguities of the text being annotated
(Aroyo et al., 2019). However, aggregating these
annotations into single ground-truth labels disre-
gards the nuances of such disagreements (Uma
et al., 2021) and even leads to disproportionate
representation of individual annotators in anno-
tated datasets (Prabhakaran et al., 2021). Here,
we explore the effect of normative social stereo-
types, as encoded in language, on the aggregated
hate labels provided in a large annotated dataset.
Annotated datasets of hate speech commonly
represent the aggregated judgments of annota-
tors rather than individual annotators’ annotation
behaviors. Therefore, rather than being impacted
by individual annotators’ self-reported social ste-
reotypes (as in Study 1), we expect aggregated
labels to be affected by normative social stereo-
types. Here, we rely on semantic representations
of social groups in pre-trained language models,
known to encode normative social stereotypes
and biases of large text corpora (Bender et al.,
2021). Figure 3 illustrates the methodology of
Study 2.
Data We analyzed the GHC (Kennedy et al.,
2022, discussed in Study 1) which includes 27,665
social-media posts labeled for hate speech content
by 18 annotators. This dataset includes 91,967
annotations in total, where each post is annotated
by at least three coders. Based on our definition of
item disagreement in Equation 1, we computed
305
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
(Nobservation = 7550), and conducted a multi-
level model, with mentioned social group tokens
as the level-1 variable and posts as the level-2
variable. We conducted two logistic regression
analyses to assess the impact of (1) the warmth
and (2) the competence of the mentioned social
group as independent variables, and with the inter-
annotator disagreement as the dependent vari-
able. The results of the two models demonstrate
that both higher warmth (β = −2.62, SE = 0.76,
p < 0.001) and higher competence (β = −5.27,
SE = 0.62, p < 0.001) scores were associated
with lower disagreement. Similar multi-level lo-
gistic regressions with the majority hate label of
the posts as the dependent variable and consider-
ing either social groups’ warmth or competence
as independent variables show that competence
predicts lower hate (β = −7.77, SE = 3.47, p =
.025), but there was no significant relationship
between perceived warmth and the hate speech
content (β = −3.74, SE = 4.05, p = 0.355). We
like to note that controlling for the frequency of
each social groups’ mentions in the dataset yields
the same results (see Supplementary Materials).
In this study, we demonstrated that social ste-
reotypes (i.e., warmth and competence), as en-
coded into language resources, are associated with
annotator disagreement in an annotated dataset
of hate speech. As in Study 1, annotators agreed
more on their judgments about social media
posts that mention stereotypically more com-
petent groups. Moreover, we observed higher
inter-annotator disagreement on social media
posts that mentioned stereotypically cold social
groups (Figure 4). While Study 1 demonstrated
novice annotators’ higher tendency for detecting
hate speech targeting stereotypically competent
groups, we found a lower likelihood of hate labels
for posts that mention stereotypically competent
social groups in this dataset. The potential reasons
for this discrepancy are: (1) while both novice and
expert annotators have been exposed to the same
definition of hate speech (Kennedy et al., 2018),
expert annotators’ training focused more on the
consequences of hate speech targeting marginal-
ized groups; moreover, the lack of variance in
expert annotators’ socio-demographic background
(mostly young, educated, liberal adults) have led
to their increased sensitivity about hate speech
directed toward specific stereotypically incompe-
tent groups; and (2) while Study 1 uses a set of
items with balanced representation for different
Figure 4: Effects of social groups’ stereotype content,
on majority hate labels, and annotators’ disagreement.
social groups, the dataset used in Study 2 in-
cludes disproportionate mentions of social groups.
Therefore,
the effect might be caused by the
higher likelihood of hateful language appearing
in GHC’s social media posts mentioning stereo-
typically less competent groups.
3 Study 3: Model Training
NLP models that are trained on human-annotated
datasets are prone to patterns of false predic-
tions associated with specific social group tokens
(Blodgett and O’Connor, 2017; Davidson et al.,
2019). For example, trained hate speech classi-
fiers may have a high probability of assigning a
hate speech label to a non-hateful post that men-
tions the word ‘‘gay.’’ Such patterns of false
predictions are known as prediction bias (Hardt
et al., 2016; Dixon et al., 2018), which impact
models’ performance on input data associated
with specific social groups. Previous research
has investigated several sources leading to pre-
diction bias, such as disparate representation of
specific social groups in the training data and
language models, or the choice of research de-
sign and machine learning algorithm (Hovy and
Prabhumoye, 2021). However, to our knowledge,
no study has evaluated prediction bias with re-
gard to the normative social stereotypes targeting
each social group. In Study 3, we investigate
whether social stereotypes influence hate speech
classifiers’ prediction bias toward those groups.
306
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
(Devlin et al., 2019) and RoBERTa (Zhuang
et al., 2021). We implemented these two classifi-
cation models using the transformers (v3.1)
library of HuggingFace (Wolf et al., 2020) and
fine-tuned both models for six epochs with a
learning rate of 10−7. The third model applies
a Support Vector Machine (SVM; Cortes and
Vapnik, 1995) with a linear kernel on Term
Frequency-Inverse Document Frequency (TF-IDF)
vector representations, implemented through the
scikit-learn (Pedregosa et al., 2011) Python
package.
Models were trained on subsets of the GHC
and their performance was evaluated on test items
mentioning different social groups. To account
for possible variations in the resulting models,
caused by selecting different subsets of the data-
set for training, we performed 100 iterations of
model training and evaluating for each classi-
fier. In each iteration, we trained the model on a
randomly selected 80% of the dataset (ntrain =
22, 132) and recorded the model predictions on
the remaining 20% of the samples (ntest =
5, 533). Then, we explored model predictions for
all iterations (nprediction = 100 × 5, 533), to cap-
ture false predictions for instances that mention
at least one social group token. By comparing
the model prediction with the majority vote for
each instance provided in GHC, we detected all
‘‘incorrect’’ predictions. For each social group,
we specifically capture the number of false-
negative (hate speech instances which are labeled
as non-hateful) and false-positive (non-hateful
instances labeled as hate speech) predictions. For
each social group token the false-positive and
false-negative ratios are calculated by dividing
the number of false predictions by the total num-
ber of posts mentioning the social group token.
Quantifying Social Stereotypes
In each analy-
sis, we considered either warmth or competence
(calculated as in Study 2) of social groups as the
independent variable to predict false-positive and
false-negative predictions as dependent variables.
Classification Results On average,
the clas-
sifiers based on BERT, RoBERTa, and SVM
achieved F1 scores of 48.22% (SD = 3%),
47.69% (SD = 3%), and 35.4% (SD = 1%),
respectively, on the test sets over the 100 itera-
tions. Since the GHC includes a varying number
of posts mentioning each social group token, the
Figure 5: The overview of Study 3. In each iteration,
the model is trained on a subset of the dataset. The false
predictions of the model are then calculated for each
social group token mentioned in test items.
We define prediction bias as erroneous predic-
tions of our text classifier model. We specifically
focus on false positives (hate-speech labels as-
signed to non-hateful instances) and false nega-
tives (non-hate-speech labels assigned to hateful
instances) (Blodgett et al., 2020).
In the two previous studies, we demonstrated
that variance in annotators’ behaviors toward hate
speech and imbalanced distribution of ground-
truth labels in datasets are both associated with
stereotypical perceptions about social groups.
Accordingly, we expect hate speech classifiers,
trained on the ground-truth labels,
to be af-
fected by stereotypes that provoke disagreements
among annotators. If that is the case, we ex-
pect
the classifier to perform less accurately
and in a biased way on social-media posts that
mention social groups with specific social stereo-
types. To detect patterns of false predictions for
specific social groups (i.e., prediction bias), we
first train several models on different subsets of
an annotated corpus of hate speech (GHC; de-
scribed in Study 1 and 2). We then evaluate the
frequency of false predictions provided for each
social group and their association with the so-
cial groups’ stereotypes. Figure 5 illustrates an
overview of this study.
Hate Speech Classifiers We implemented three
hate speech classifiers; the first two models are
based on pre-trained language models, BERT
307
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
predictions (nprediction = 553, 300) include a
varying number of items for each social group
token (M = 2,284.66, M dn = 797.50, SD =
3,269.20). ‘‘White’’ as the most frequent social
group token appears in 16,155 of the predictions
and ‘‘non-binary’’ is the least frequent social
group token with only 13 observations. Since so-
cial group tokens have varying distributions in
the dataset, we considered the ratios of false pre-
dictions (rather than frequencies) in all regres-
sion models by adding the log-transform of the
number of test samples for each social group
token as the offset.
‘‘Buddhist.’’
Analysis of Results The average false-positive
ratio of social group tokens in the BERT-classifier
was 0.58 (SD = 0.24), with a maximum of 1.00
false-positive ratio for several social groups, in-
cluding ‘‘bisexual’’, and the minimum of 0.03
false-positive ratio for
In other
words, BERT-classifiers always predicted incor-
rect hate speech labels for non-hateful social-
media posts mentioning ‘‘bisexuals’’ while rarely
making those mistakes for posts mentioning
‘‘Buddhists’’. The average false-negative ratio of
social group tokens in the BERT-classifier was
0.12 (SD = 0.11), with a maximum of 0.49
false-negative ratio associated with ‘‘homosex-
ual’’ and the minimum of 0.0 false-negative ratio
for several social groups including ‘‘Latino.’’ In
other words, BERT-classifiers predicted incorrect
non-hateful labels for social-media post mention-
ing ‘‘homosexuals’’ while hardly making those
mistakes for posts mentioning ‘‘Latino’’. These
statistics are consistent with observations of pre-
vious findings (Davidson et al., 2017; Kwok and
Wang, 2013; Dixon et al., 2018; Park et al., 2018),
which identify false-positive errors as the more
critical issue with hate speech classifiers.
For each classifier, we assess the number
of false-positive and false-negative hate speech
predictions for social-media posts that mention
each social group. For analyzing each classifier,
two Poisson models were created, consider-
ing false-positive predictions as the dependent
variable and social groups’ (1) warmth or (2) com-
petence, calculated from a pre-trained language
model (see Study 2) as the independent variable.
The same settings were considered in two other
Poisson models to assess false-negative predic-
tions as the dependent variable, and either warmth
or competence as the independent variable.
False Positive
C
W
False Negative
W
C
BERT
RoBERTa −0.04** −0.15**
SVM
−0.09** −0.23** −0.04** −0.10**
0.05*
−0.09**
0.02
−0.05** −0.09** −0.01
Table 1: Associations of erroneous predictions
(false positives and false negatives) and so-
cial groups’ warmth (W) and competence (C)
stereotypes in predictions of three classifiers. **
and * represent p-values less than .001 and .05,
respectively.
Table 1 reports the association between so-
cial groups’ warmth and competence stereotypes
with the false hate speech labels predicted by
the models. The results indicate that the number
of false-positive predictions is negatively associ-
ated with the social groups’ language-embedded
warmth and competence scores in all three mod-
els. In other words, texts that mentions social
groups stereotyped as cold and incompetent are
more likely to be misclassified as containing hate
speech; for instance, in the BERT-classifier a one
point increase in the social groups warmth and
competence is, respectively, associated with 8.4%
and 20.3% decrease in model’s false-positive error
ratios. The number of false-negative predictions
is also significantly associated with the social
groups’ competence scores; however, this asso-
ciation had varying directions among the three
models. BERT and SVM classifiers are more
likely to misclassify instances as not containing
hate speech when texts mention stereotypically
incompetent social groups; such that one point
increase in competence is associated with 9.8%
decrease in BERT model’s false-negative error
ratio. Whereas false-negative predictions of the
RoBERTa model is more likely for text men-
tioning stereotypically competent social groups.
The discrepancy in the association of warmth and
competence stereotypes and false-negative errors
calls for further investigation. Figure 6 depicts
the associations of the two stereotype dimen-
sions with the proportions of false-positive and
false-negative predictions of the BERT classifier
for social groups.
In summary, this study demonstrates that erro-
neous predictions of hate speech classifiers are
associated with the normative stereotypes re-
garding the social groups mentioned in text.
308
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
media posts about those social groups. Our find-
ings indicate that for novice annotators judging
social groups as competent is associated with a
higher tendency toward detecting hate and lower
disagreement with other annotators. We reasoned
that novice annotators prioritize protecting the
groups they perceive as warm and competent.
These results can be interpreted based on the Be-
haviors from Intergroup Affect and Stereotypes
framework (BIAS; Cuddy et al., 2007): groups
judged as competent elicit passive facilitation (i.e.,
obligatory association), whereas those judged as
lacking competence elicit passive harm (i.e., ig-
noring). Here, novice annotators might tend to
‘‘ignore’’ social groups judged to be incompe-
tent and not assign ‘‘hate speech’’ labels to in-
flammatory posts attacking these social groups.
However, Study 1’s results may not uncover
the pattern of annotation biases in hate speech
datasets as data curation efforts rely on annotator
pools with imbalanced representation of different
socio-demographic groups (Posch et al., 2018) and
data selection varies among different datasets. In
Study 2, we examined the role of social stereotypes
in the aggregation process, where expert annota-
tors’ disagreements are discarded to create a large
dataset containing the ground-truth hate-speech
labels. We demonstrated that, similar to Study 1,
texts that included groups stereotyped to be warm
and competent were highly agreed upon. How-
ever, unlike Study 1, posts mentioning groups
stereotyped as incompetent are more frequently
marked as hate speech by the aggregated labels.
In other words, novice annotators tend to focus
on protecting groups they perceive as competent;
however, the majority vote of expert annotators
tend to focus on common targets of hate in the
corpus. We noted two potential reasons for this
disparity (1) Novice and expert annotators vary
in their annotation behaviors; in many cases, hate
speech datasets are labeled by expert annotators
who are thoroughly trained for this specific task
(Patton et al., 2019), and have specific experi-
ences that affect their perception of online hate
(Talat, 2016). GHC annotators were undergrad-
uate psychologist research assistants trained by
first reading a typology and coding manual for
studying hate-based rhetoric and then passing a
curated test of about thirty messages designed for
assessing their understanding of the annotation
task (Kennedy et al., 2022). Therefore, their rel-
atively higher familiarity with and experience in
Figure 6: Social groups’ higher stereotypical compe-
tence and warmth is associated with lower false positive
and negative predictions in hate speech detection.
Particularly, the results indicate that documents
mentioning stereotypically colder and less compe-
tent social groups, which lead to higher disagree-
ment among expert annotators based on Study 2,
drive higher error rates in hate speech classi-
fiers. This pattern of high false predictions (both
false-positives and false-negatives)
for social
groups stereotyped as cold and incompetent im-
plies that prediction bias in hate speech classifiers
is associated with social stereotypes, and resem-
bles normative social biases that we documented
in the previous studies.
4 Discussion
Here, we integrate theory-driven and data-driven
approaches (Wagner et al., 2021) to investigate
human annotators’ and normative social stereo-
types as a source of bias in hate speech data-
sets and classifiers. In three studies, we combine
social psychological frameworks and computa-
tional methods to make theory-driven predictions
about hate-speech-annotation behavior and em-
pirically test the sources of bias in hate speech
classifiers. Overall, we find that hate speech an-
notation behaviors, often assumed to be objective,
are impacted by social stereotypes, and that this
in turn adversely influences automated content
moderation.
In Study 1, we investigated the association
between participants’ self-reported social stereo-
types against 8 different social groups, and their
annotation behavior on a small subset of social-
309
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
annotating hate speech, compared to annotators
in Study 1, led to different annotation behaviors.
Moreover, dataset annotators are not usually rep-
resentative of the exact population that interacts
with social media content. As pointed out by D´ıaz
et al. (2022), understanding the socio-cultural fac-
tors of an annotator pool can shed light on the
disparity of our results. In our case, identities and
lived experiences can significantly vary between
participants in Study 1 and GHC’s annotators in
Study 2, which impacts how annotation ques-
tions are interpreted and responded to. (2) Social
groups with specific stereotypes have imbalanced
presence in hate speech datasets; while in Study 1,
we collect a balanced set of items with equal rep-
resentation for each of the 8 social groups, social
media posts disproportionately include mentions
of different social groups, and the frequency of
each social group being targeted depends on mul-
tiple social and contextual factors.
To empirically demonstrate the effect of social
stereotypes on supervised hate speech classifiers,
in Study 3, we evaluated the performance and
biased predictions of such models when trained
on an annotated dataset. We used the ratio of
incorrect predictions to operationalize the clas-
sifiers’ unintended bias in assessing hate speech
toward specific groups (Hardt et al., 2016). Study
3’s findings suggested that social stereotypes of
a mentioned group, as captured in large language
models, are significantly associated with biased
classification of hate speech such that more false-
positive predictions are generated for documents
that mention groups that are stereotyped to be
cold and incompetent. However, we did not find
consistent trends in associations between social
groups’ warmth and competence stereotypes and
false-negative predictions among different mod-
els. These results demonstrate that false-positive
predictions are more frequent for the same social
groups that evoked more disagreements between
annotators in Study 2. Similar to Davani et al.
(2022), these findings challenge supervised learn-
ing approaches that only consider the majority
vote for training a hate speech classifier and dis-
pose of the annotation biases reflected in inter-
annotator disagreements.
It should be noted that while Study 1 assesses so-
cial stereotypes as reported by novice annotators,
Studies 2 and 3 rely on a semantic representa-
tion of such stereotypes. Since previous work on
language representation have shown that semantic
representations encode socially embedded biases,
in Studies 2 and 3 we referred to the construct
under study as normative social stereotypes. Our
comparison of results demonstrated that novice
annotators’ self-reported social stereotypes im-
pact their annotation behaviors, and the annotated
datasets and hate speech classifiers are prone to
being affected by normative stereotypes.
Our work is limited to the English language,
a single dataset of hate speech, and participants
from the US. Given that the increase in hate
speech is not limited to the US, it is important
to extend our findings in terms of research par-
ticipants and language resources. Moreover, we
applied SCM to quantify social stereotypes, but
other novel theoretical frameworks such as the
Agent-Beliefs-Communion model (Koch et al.,
2016) can be applied in the future to uncover
other sources of bias.
5 Related Work
Measuring Annotator Bias Annotators are bi-
ased in their interpretations of subjective language
understanding tasks (Aroyo et al., 2019; Talat
et al., 2021). Annotators’ sensitivity to toxic lan-
guage can vary based on their expertise (Talat,
2016), lived experiences (Patton et al., 2019), and
demographics (e.g., gender, race, and political
orientation) (Cowan et al., 2002; Norton and
Sommers, 2011; Carter and Murphy, 2015;
Prabhakaran et al., 2021; Jiang et al., 2021).
Sap et al. (2022) discovered associations between
annotators’ racist beliefs and their perceptions of
toxicity in anti-Black messages and text written
in African American English. Compared to pre-
vious efforts, our research takes a more general
approach to modeling annotators’ biases, which
is not limited to specific targets of hate.
Recent research efforts argue that annotators’
disagreements should not be treated solely as
noise in data (Pavlick and Kwiatkowski, 2019)
and call for alternative approaches for consid-
ering annotators as independent sources for in-
forming the modeling process in subjective tasks
(Prabhakaran et al., 2021). Such efforts tend to im-
prove data collection (Vidgen et al., 2021; Rottger
et al., 2022) and the modeling process in various
tasks, such as detecting sarcasm (Rajadesingan
et al., 2015), humor (Gultchin et al., 2019), senti-
ment (Gong et al., 2017), and hate speech (Koco´n
et al., 2021). For instance, Davani et al. (2022)
310
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
introduced a method for modeling individual an-
notators’ behaviors rather than their majority vote.
In another work, Akhtar et al. (2021) clustered
annotators into groups with high internal agree-
ment (similarly explored by Wich et al., 2020)
and redefined the task as modeling the aggre-
gated label of each group. Our findings especially
help such efforts by providing a framework for
incorporating annotators’ biases into hate speech
classifiers.
Measuring Hate Speech Detection Bias When
propagated into the modeling process, biases in
the annotated hate speech datasets cause group-
based biases in predictions (Sap et al., 2019) and
lack of robustness in results (Geva et al., 2019;
Arhin et al., 2021). Specifically, previous research
has shed light on unintended biases (Dixon et al.,
2018), which are generally defined as systemic
differences in performance for different demo-
graphic groups, potentially compounding existing
challenges to fairness in society at large (Borkan
et al., 2019). While a significant body of work has
been dedicated to mitigating unintended biases
in hate speech (and abusive language) classifi-
cation (Vaidya et al., 2020; Ahmed et al., 2022;
Garg et al., 2019; Nozza et al., 2019; Badjatiya
et al., 2019; Park et al., 2018; Mozafari et al.,
2020; Xia et al., 2020; Kennedy et al., 2020;
Mostafazadeh Davani et al., 2021; Chuang et al.,
2021), the choice of the exact bias metrics is not
consistent within all these studies. As demon-
strated by Czarnowska et al. (2021), various bias
metrics can be considered as different parametri-
zations of a generalized metric. In hate speech
detection in particular, disproportionate false pre-
dictions, especially false positive predictions, for
marginalized social groups have often been con-
sidered as an indicator of unintended bias in the
model. This is due to the fact that hate speech, by
definition, involves a social group as the target of
hate, and the disproportionate mentions of specific
social groups in hateful social media content have
led to imbalance datasets and biased models.
Measuring Social Stereotypes The Stereotype
Content Model (SCM; Fiske et al., 2002) sug-
gests that to determine whether other people are
threats or allies, individuals make prompt assess-
ments about their warmth (good vs. ill intentions)
and competence (ability vs. inability to act on
intentions). Koch et al. (2016) proposed to fill
in an empirical gap in SCM by introducing the
ABC model of stereotype content. Based on this
model, people organize social groups primarily
based on their (A) agency (competence in SCM),
and (B) conservative-progressive beliefs. They
did not find (C) communion (warmth in SCM)
as a dimension by its own, but rather as an
emergent quality in the other two dimensions.
Zou and Cheryan (2017) proposed that racial and
ethnic minority groups are disadvantaged along
two distinct dimensions of perceived inferiority
and perceived cultural foreignness, which can
explain the patterns of racial prejudice.
Among different social psychological frame-
works for social stereotypes, we relied on SCM
in this research, as it has been shown helpful for
predicting emotional and behavioral reactions to
outgroups. For instance, the Behaviors from Inter-
group Affect and Stereotypes framework (BIAS;
Cuddy et al., 2007), an extension of the SCM, ex-
plains how stereotypes shape behavioral tenden-
cies toward groups and individuals (Cuddy et al.,
2008). Based on this theoretical framework, per-
ceived warmth predicts active behaviors, while
perceived competence predicts passive behaviors
toward outgroups. Individuals tend to express ad-
miration for social groups stereotyped as warm
and competent and facilitate them both actively
(by protecting them) and passively (by cooper-
ating with them). In contrast, people tend to ex-
press hate toward social groups stereotyped as
cold and incompetent and harm them actively (by
attacking) and passively (by excluding).
6 Conclusion
Our findings suggest that hate speech classifiers
trained on human annotations will acquire partic-
ular prediction biases toward historically margin-
alized groups. Our results have two specific and
direct implications: First, one way to decrease un-
intended bias in classifiers is to study and diver-
sify annotation teams, so that annotators come
from different ideological and social backgrounds;
by doing so, coders may not agree on a label to
(over)protect an ingroup or ‘‘ally’’ groups (lead-
ing to false positives), or to ignore actually hateful
content targeting ‘‘blameworthy’’ groups (leading
to false negatives). Second, supervised learning
approaches may benefit from modeling annota-
tion biases, which are reflected in inter-annotator
disagreements, rather than the current practice,
311
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
which is to treat them as unexplained noise in
human judgment, to be disposed of through an-
notation aggregation. Our work is an example of
how well-established social scientific theories can
be applied to explain the novel interactions be-
tween algorithms and people. Large amounts of
data that are being constantly recorded in ever-
changing socio-technical environments call for
integrating social scientific theories for assessing
the novel technologies and their associated so-
cial impacts.
A Supplementary Materials
Study 1 – Test Items The three test items are:
(1) I want to see Sandirians raped to death col-
lectively. (2) Now if we could only get a few
jew heads on pikes that would be great. (3) Over
99% of Syrian refugees admitted to UK in 2017
were Muslims. We expect annotators to label (1)
and (2) as hate speech and label (3) as not hate
speech.
Study 1 – Analysis of All Annotators We repli-
cate the results of Study 1, on the whole set
of participants (N = 1,228). The result shows
that a higher number of items are categorized as
hate speech when participants perceive that so-
cial group as high on competence (β = 0.02,
SE = 0.005, p < .001). However, warmth scores
were not significantly associated with the number
of hate-speech labels (β = 0.01, SE = 0.006,
p = .286). Moreover, participants demonstrate
higher tendency (to assign hate speech labels) on
items that mention a social group they perceive as
(β = 0.04, SE = 0.010,
highly competent
p < .001). Warmth scores were only marginally
associated with participants’ tendency scores (β =
0.02, SE = 0.010, p = 0.098). Lastly, participants
disagreed more on items that mention a social
group perceived as incompetent (β = −0.17,
SE = 0.034, p <.001). Contrary to the origi-
nal results, warmth scores were also significantly
associated with the odds of disagreement (β =
0.07, SE = 0.036, p = .044).
Study 1 and 2 – Stereotypes Table 2 reports the
calculated stereotype scores for each social group.
Study 2 assesses over 63 social groups; the
calculated warmth score varies from 0.01 to 0.19
(mean = 0.14, sd = 0.03), and competence varies
from −0.03 to 0.22 (mean = 0.14, sd = 0.04).
Group
Immigrant
Muslim
Communist
Liberal
Black
Gay
Jewish
Woman
Study 1
C
6.8 (1.6)
7.0 (1.7)
5.8 (2.0)
6.7 (2.0)
7.0 (1.7)
7.3 (1.5)
7.7 (1.3)
7.6 (1.3)
W
7.2 (1.4)
6.6 (1.8)
5.1 (2.0)
6.6 (1.9)
6.9 (1.6)
7.5 (1.4)
7.3 (1.4)
7.5 (1.2)
Study 2
C W
5.0
5.0
5.1
4.9
5.0
5.0
5.1
5.2
4.7
4.8
5.1
4.9
5.0
4.9
5.1
5.2
Table 2: Perceived warmth (W) and competence
(C) scores varying from 1 (most negative trait) to
8 (most positive trait). Study 1 columns represent
the average and standard deviation of participants’
responses. Study 2’s values are scaled from [−1,
1] to [1, 8]. The correlation of perceived compe-
tence and warmth score within the two studies are
−0.07 and 0.09, respectively.
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Figure 7: The distribution of social groups on the
warmth-competence space based the calculated scores
used in Study 2.
Figure 7 plots the social groups on the warmth and
competence dimensions calculated in Study 2.
Study 2 – Frequency as a Control Variable
After adding social groups’ frequency as a control
variable, both higher warmth (β = −2.28, SE =
0.76, p < 0.01) and competence (β = −5.32,
SE = 0.62, p < 0.001) scores were associated
with lower disagreement. Competence predicts
lower hate (β = −7.96, SE = 3.71, p = .032),
but there was no significant relationship between
perceived warmth and the hate speech content
(β = −2.95, SE = 3.89, p = .448).
312
Acknowledgments
We would like to thank Nils Karl Reimer,
Vinodkumar Prabhakaran, Stephen Read,
the
anonymous reviewers, and the action editor for
their suggestions and feedback.
References
Zo Ahmed, Bertie Vidgen, and Scott A. Hale.
2022. Tackling racial bias in automated on-
line hate detection: Towards fair and accur-
ate detection of hateful users with geometric
deep learning. EPJ Data Science, 11(1):8.
https://doi.org/10.1140/epjds/s13688
-022-00319-9
Sohail Akhtar, Valerio Basile, and Viviana Patti.
2021. Whose opinions matter? Perspective-
aware models to identify opinions of hate
speech victims in abusive language detection.
arXiv preprint arXiv:2106.15896.
Kofi Arhin,
Ioana Baldini, Dennis Wei,
Karthikeyan Natesan Ramamurthy,
and
Moninder Singh. 2021. Ground-truth, whose
truth?–examining the challenges with an-
notating toxic text datasets. arXiv preprint
arXiv:2112.03529.
Lora Aroyo, Lucas Dixon, Nithum Thain, Olivia
Redfield, and Rachel Rosen. 2019. Crowd-
sourcing subjective tasks: The case study of
understanding toxicity in online discussions.
In Companion Proceedings of The 2019 World
Wide Web Conference, pages 1100–1105.
https://doi.org/10.1145/3308560
.3317083
Pinkesh Badjatiya, Manish Gupta, and Vasudeva
Varma. 2019. Stereotypical bias removal for
hate speech detection task using knowledge-
based generalizations. In The World Wide Web
Conference, pages 49–59.
Emily M. Bender, Timnit Gebru, Angelina
McMillan-Major, and Shmargaret Shmitchell.
2021. On the dangers of stochastic parrots:
Can language models be too big? In Pro-
ceedings of
the 2021 ACM Conference on
Fairness, Accountability, and Transparency,
pages 610–623. https://doi.org/10.1145
/3442188.3445922
Su Lin Blodgett, Solon Barocas, Hal Daum´e III,
and Hanna Wallach. 2020. Language (tech-
nology) is power: A critical survey of ‘‘bias’’
in NLP. In Proceedings of the 58th Annual
Meeting of the Association for Computational
Linguistics, pages 5454–5476, Online. Associ-
ation for Computational Linguistics. https://
doi.org/10.18653/v1/2020.acl-main.485
Su Lin Blodgett and Brendan O’Connor. 2017.
Racial disparity in natural language processing:
A case study of social media African Ameri-
can English. In Fairness, Accountability, and
Transparency in Machine Learning (FAT/ML)
Workshop, KDD.
Tolga Bolukbasi, Kai-Wei Chang, James Y. Zou,
Venkatesh Saligrama, and Adam T. Kalai.
2016. Man is to computer programmer as
woman is to homemaker? Debiasing word em-
beddings. In Advances in Neural Information
Processing Systems, pages 4349–4357.
Daniel Borkan, Lucas Dixon, Jeffrey Sorensen,
Nithum Thain, and Lucy Vasserman. 2019. Nu-
anced metrics for measuring unintended bias
with real data for text classification. In Com-
panion Proceedings of the 2019 World Wide
Web Conference, pages 491–500.
Aylin Caliskan, Joanna J. Bryson, and Arvind
Narayanan. 2017. Semantics derived automati-
cally from language corpora contain human-like
biases. Science, 356(6334):183–186. https://
doi.org/10.1126/science.aal4230, PubMed:
28408601
Evelyn R. Carter and Mary C. Murphy. 2015.
Group-based differences in perceptions of
racism: What counts,
to whom, and why?
Social and Personality Psychology Com-
pass, 9(6):269–280. https://doi.org/10
.1111/spc3.12181
Tessa E. S. Charlesworth, Victor Yang, Thomas
C. Mann, Benedek Kurdi, and Mahzarin
R. Banaji. 2021. Gender stereotypes in natu-
ral language: Word embeddings show robust
consistency across child and adult language
corpora of more than 65 million words. Psy-
chological Science, 32:218–240. https://doi
.org/10.1177/0956797620963619, PubMed:
33400629
Yung-Sung Chuang, Mingye Gao, Hongyin Luo,
James Glass, Hung-yi Lee, Yun-Nung Chen,
313
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
and Shang-Wen Li. 2021. Mitigating biases
in toxic language detection through invari-
ant rationalization. In Proceedings of the 5th
Workshop on Online Abuse and Harms (WOAH
2021), pages 114–120, Online. Association for
Computational Linguistics.
Corinna Cortes and Vladimir Vapnik. 1995.
Support-vector networks. Machine Learning,
20(3):273–297. https://doi.org/10.1007
/BF00994018
Gloria Cowan, Miriam Resendez, Elizabeth
Marshall, and Ryan Quist. 2002. Hate speech
and constitutional protection: Priming values
of equality and freedom. Journal of Social
Issues, 58(2):247–263. https://doi.org
/10.1111/1540-4560.00259
Kate Crawford. 2017. The trouble with bias. In
Conference on Neural Information Processing
Systems, invited speaker.
Amy J. C. Cuddy, Susan T. Fiske, and Peter
Glick. 2007. The bias map: Behaviors from
intergroup affect and stereotypes. Journal of
Personality and Social Psychology, 92(4):
631–648. https://doi.org/10.1037/0022
-3514.92.4.631, PubMed: 17469949
Amy J. C. Cuddy, Susan T. Fiske, and Peter
Glick. 2008. Warmth and competence as uni-
versal dimensions of social perception: The
stereotype content model and the bias map.
Advances in Experimental Social Psychology,
40:61–149. https://doi.org/10.1016
/S0065-2601(07)00002-0
Paula Czarnowska, Yogarshi Vyas, and Kashif
Shah. 2021. Quantifying social biases in NLP:
A generalization and empirical comparison
of extrinsic fairness metrics. Transactions of
the Association for Computational Linguis-
tics, 9:1249–1267. https://doi.org/10
.1162/tacl_a_00425
Aida Mostafazadeh Davani, Mark D´ıaz, and
Vinodkumar Prabhakaran. 2022. Dealing with
disagreements: Looking beyond the majority
vote in subjective annotations. Transactions of
the Association for Computational Linguistics,
10:92–110. https://doi.org/10.1162
/tacl_a_00449
Thomas Davidson, Debasmita Bhattacharya,
and Ingmar Weber. 2019. Racial bias in
hate speech and abusive language detection
datasets. In Proceedings of the Third Workshop
on Abusive Language Online, pages 25–35,
Florence,
Italy. Association for Computa-
tional Linguistics. https://doi.org/10
.18653/v1/W19-3504
Thomas Davidson, Dana Warmsley, Michael
Macy, and Ingmar Weber. 2017. Automated
hate speech detection and the problem of offen-
sive language. In Eleventh international aaai
conference on web and social media.
J. Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training
of deep bidirectional transformers for language
understanding. In NAACL-HLT.
Mark D´ıaz,
Ian Kivlichan, Rachel Rosen,
Dylan Baker, Razvan Amironesei, Vinodkumar
Prabhakaran, and Emily Denton. 2022. Crowd-
worksheets: Accounting for
individual and
collective identities underlying crowdsourced
dataset annotation. In 2022 ACM Conference
on Fairness, Accountability, and Transparency,
pages 2342–2351. https://doi.org/10
.1145/3531146.3534647
Lucas Dixon, John Li, Jeffrey Sorensen, Nithum
Thain, and Lucy Vasserman. 2018. Measur-
ing and mitigating unintended bias in text
the 2018
classification.
AAAI/ACM Conference on AI, Ethics, and So-
ciety, pages 67–73. https://doi.org/10
.1145/3278721.3278729
In Proceedings of
Michael Feldman, Sorelle A. Friedler, John
Moeller, Carlos Scheidegger, and Suresh
Venkatasubramanian. 2015. Certifying and re-
moving disparate impact. In proceedings of
the 21th ACM SIGKDD international confer-
ence on knowledge discovery and data mining,
pages 259–268.
Susan T. Fiske, Amy J. C. Cuddy, P. Glick,
and J. Xu. 2002. A model of (often mixed)
stereotype content: Competence and warmth
respectively follow from perceived status and
competition. Journal of Personality and Social
Psychology, 82(6):878. https://doi.org
/10.1037/0022-3514.82.6.878
Joseph L. Fleiss. 1971. Measuring nominal
scale agreement among many raters. Psycho-
logical bulletin, 76(5):378. https://doi
.org/10.1037/h0031619
314
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Gavin Gaffney. 2018. Pushshift gab corpus.
https://files.pushshift.io/gab/.
Accessed: 2019-5-23.
Social Psychology Bulletin, 31(10):1369–1385.
https://doi.org/10.1177/0146167205275613,
PubMed: 16143669
Nikhil Garg, Londa Schiebinger, Dan Jurafsky,
and James Zou. 2018. Word embeddings
quantify 100 years of gender and ethnic stereo-
types. Proceedings of the National Academy
of Sciences, 115(16):E3635–E3644. https://
doi.org/10.1073/pnas.1720347115, PubMed:
29615513
Sahaj Garg, Vincent Perot, Nicole Limtiaco,
Ankur Taly, Ed H. Chi, and Alex Beutel.
2019. Counterfactual fairness in text classifi-
cation through robustness. In Proceedings of
the 2019 AAAI/ACM Conference on AI, Ethics,
and Society, pages 219–226. https://doi
.org/10.1145/3306618.3317950
Mor Geva, Yoav Goldberg, and Jonathan Berant.
the
2019. Are we modeling the task or
annotator? An investigation of annotator bias
in natural language understanding datasets. In
Proceedings of the 2019 Conference on Empiri-
cal Methods in Natural Language Processing
and the 9th International Joint Conference
on Natural Language Processing (EMNLP-
IJCNLP), pages 1161–1166, Hong Kong, China.
Association for Computational Linguistics.
https://doi.org/10.18653/v1/D19
-1107
Lin Gong, Benjamin Haines, and Hongning
Wang. 2017. Clustered model adaption for
personalized sentiment analysis. In Proceed-
ings of the 26th International Conference on
World Wide Web, pages 937–946.
Limor Gultchin, Genevieve Patterson, Nancy
Baym, Nathaniel Swinger, and Adam Kalai.
2019. Humor in word embeddings: Cocka-
mamie gobbledegook for nincompoops.
In
International Conference on Machine Learn-
ing, pages 2474–2483. PMLR.
Moritz Hardt, Eric Price, and Nati Srebro. 2016.
Equality of opportunity in supervised learning.
In Advances in Neural Information Processing
Systems, pages 3315–3323.
Wilhelm Hofmann, Bertram Gawronski, Tobias
Gschwendner, Huy Le, and Manfred Schmitt.
2005. A meta-analysis on the correlation
between the implicit association test and ex-
plicit self-report measures. Personality and
Dirk Hovy, Taylor Berg-Kirkpatrick, Ashish
Vaswani, and Eduard Hovy. 2013. Learning
whom to trust with mace. In Proceedings of
the 2013 Conference of the North American
Chapter of the Association for Computational
Linguistics: Human Language Technologies,
pages 1120–1130.
Dirk Hovy and Shrimai Prabhumoye. 2021.
Five sources of bias in natural
language
processing. Language and Linguistics Com-
pass, 15(8):e12432. https://doi.org/10
.1111/lnc3.12432, PubMed: 35864931
L. Rowell Huesmann, Eric F. Dubow, Paul
Boxer, Violet Souweidane, and Jeremy Ginges.
2012. Foreign wars and domestic prejudice:
How media exposure to the Israeli-Palestinian
conflict predicts ethnic stereotyping by Jewish
and Arab American adolescents. Journal of
Research on Adolescence, 22(3):556–570.
https://doi.org/10.1111/j.1532-7795
.2012.00785.x, PubMed: 23243381
Ben Hutchinson, Vinodkumar Prabhakaran, Emily
Denton, Kellie Webster, Yu Zhong, and
Stephen Denuyl. 2020. Social biases in NLP
models as barriers for persons with disabili-
ties. In Proceedings of the 58th Annual Meeting
of the Association for Computational Linguis-
tics, pages 5491–5501, Online. Association
for Computational Linguistics.
Jialun Aaron Jiang, Morgan Klaus Scheuerman,
Casey Fiesler, and Jed R. Brubaker. 2021.
Understanding international perceptions of the
severity of harmful content online. PloS One,
16(8):e0256762. https://doi.org/10.1371
/journal.pone.0256762, PubMed: 34449815
Brendan Kennedy, Mohammad Atari, Aida
Mostafazadeh Davani, Leigh Yeh, Ali Omrani,
Yehsong Kim, Kris Coombs, Shreya Havaldar,
Gwenyth Portillo-Wightman, Elaine Gonzalez,
Joe Hoover, Aida Azatian, Alyzeh Hussain,
Austin Lara, Gabriel Cardenas, Adam Omary,
Christina Park, Xin Wang, Clarisa Wijaya,
Yong Zhang, Beth Meyerowitz, and Morteza
Dehghani. 2022.
Introducing the gab hate
corpus: Defining and applying hate-based rhet-
oric to social media posts at scale. Language
Resources and Evaluation, 56(1):79–108.
315
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
https://doi.org/10.1007/s10579-021
-09569-x
Natural Language Processing, volume 2016,
page 648. NIH Public Access.
Brendan Kennedy, Xisen Jin, Aida Mostafazadeh
Davani, Morteza Dehghani, and Xiang Ren.
2020. Contextualizing hate speech classifiers
with post-hoc explanation. In Proceedings of
the 58th Annual Meeting of the Association for
Computational Linguistics, pages 5435–5442,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653
/v1/2020.acl-main.483
Brendan Kennedy, Drew Kogon, Kris Coombs,
Joseph Hoover, Christina Park, Gwenyth
Portillo-Wightman, Aida Mostafazadeh Davani,
Mohammad Atari, and Morteza Dehghani. 2018.
A typology and coding manual for the study of
hate-based rhetoric. PsyArXiv. July, 18.
Svetlana Kiritchenko,
Isar Nejadgholi,
and
Kathleen C. Fraser. 2021. Confronting abusive
language online: A survey from the ethical and
human rights perspective. Journal of Artificial
Intelligence Research, 71:431–478. https://
doi.org/10.1613/jair.1.12590
Alex Koch, Roland Imhoff, Ron Dotsch, Christian
Unkelbach, and Hans Alves. 2016. The abc of
stereotypes about groups: Agency/socioeconomic
success, conservative–progressive beliefs, and
communion. Journal of Personality and So-
cial Psychology, 110(5):675. https://doi
.org/10.1037/pspa0000046, PubMed:
27176773
Jan Koco´n, Marcin Gruza, Julita Bielaniewicz,
Damian Grimling, Kamil Kanclerz, Piotr
Miłkowski, and Przemysław Kazienko. 2021.
Learning personal human biases and rep-
resentations for subjective tasks in natural
language processing. In 2021 IEEE Interna-
tional Conference on Data Mining (ICDM),
pages 1168–1173. IEEE.
Irene Kwok and Yuzhou Wang. 2013. Locate the
hate: Detecting tweets against blacks. In Pro-
ceedings of the AAAI Conference on Artificial
Intelligence, volume 27. https://doi.org
/10.1609/aaai.v27i1.8539
John P. Lalor, Hao Wu, and Hong Yu. 2016. Build-
ing an evaluation scale using item response
theory. In Proceedings of the Conference on
Empirical Methods in Natural Language Pro-
cessing. Conference on Empirical Methods in
316
Thomas Manzini, Lim Yao Chong, Alan W.
Black, and Yulia Tsvetkov. 2019. Black is
to criminal as caucasian is to police: De-
tecting and removing multiclass bias in word
embeddings. In Proceedings of the 2019 Con-
ference of
the North American Chapter of
the Association for Computational Linguis-
tics: Human Language Technologies, Volume 1
(Long and Short Papers), pages 615–621,
Minneapolis, Minnesota. Association for Com-
putational Linguistics. https://doi.org
/10.18653/v1/N19-1062
Melissa D. McCradden, Shalmali Joshi, James A.
Anderson, Mjaye Mazwi, Anna Goldenberg,
and Randi Zlotnik Shaul. 2020. Patient
safety and quality improvement: Ethical prin-
ciples for a regulatory approach to bias in
healthcare machine learning. Journal of the
American Medical
Informatics Association,
27(12):2024–2027. https://doi.org/10
.1093/jamia/ocaa085, PubMed: 32585698
Ninareh Mehrabi, Fred Morstatter, Nripsuta
Saxena, Kristina Lerman, and Aram Galstyan.
2021. A survey on bias and fairness in machine
learning. ACM Computing Surveys (CSUR),
54(6):1–35. https://doi.org/10.1145
/3457607
Aida Mostafazadeh Davani, Ali Omrani, Brendan
Kennedy, Mohammad Atari, Xiang Ren, and
Morteza Dehghani. 2021. Improving counter-
factual generation for fair hate speech de-
tection. In Proceedings of the 5th Workshop
on Online Abuse and Harms (WOAH 2021),
pages 92–101, Online. Association for Compu-
tational Linguistics.
Marzieh Mozafari, Reza Farahbakhsh, and No¨el
Crespi. 2020. Hate speech detection and racial
bias mitigation in social media based on bert
model. PloS ONE, 15(8):e0237861. https://
doi.org/10.1371/journal.pone.0237861,
PubMed: 32853205
Michael I. Norton and Samuel R. Sommers. 2011.
Whites see racism as a zero-sum game that they
are now losing. Perspectives on Psychological
Science, 6(3):215–218. https://doi.org/10
.1177/1745691611406922, PubMed: 26168512
Debora Nozza, Claudia Volpetti, and Elisabetta
Fersini. 2019. Unintended bias in misogyny
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
In IEEE/WIC/ACM International
detection.
Conference on Web Intelligence, pages 149–155.
https://doi.org/10.1145/3350546
.3352512
Ziad Obermeyer, Brian Powers, Christine Vogeli,
and Sendhil Mullainathan. 2019. Dissect-
ing racial bias
in an algorithm used to
manage the health of populations. Science,
366(6464):447–453. https://doi.org/10
.1126/science.aax2342, PubMed: 31649194
Ji Ho Park, Jamin Shin, and Pascale Fung. 2018.
Reducing gender bias in abusive language de-
tection. In Proceedings of the 2018 Conference
on Empirical Methods in Natural Language
Processing, Brussels, Belgium. Association for
Computational Linguistics.
Desmond Patton, Philipp Blandfort, William
Frey, Michael Gaskell, and Svebor Karaman.
2019. Annotating social media data from vul-
nerable populations: Evaluating disagreement
between domain experts and graduate stu-
dent annotators. In Proceedings of the 52nd
Hawaii International Conference on System
Sciences. https://doi.org/10.24251
/HICSS.2019.260
Ellie Pavlick and Tom Kwiatkowski. 2019. Inher-
ent disagreements in human textual inferences.
Transactions of the Association for Computa-
tional Linguistics, 7:677–694. https://doi
.org/10.1162/tacl_a_00293
Fabian Pedregosa, Ga¨el Varoquaux, Alexandre
Gramfort, Vincent Michel, Bertrand Thirion,
Peter
Olivier Grisel, Mathieu Blondel,
Prettenhofer, Ron Weiss, Vincent Dubourg,
Jake Vanderplas, Alexandre Passos, and David
Cournapeau. 2011. Scikit-learn: Machine learn-
ing in Python. Journal of Machine Learning
Research, 12:2825–2830.
Jeffrey
Socher,
Pennington, Richard
and
Christopher D. Manning. 2014. GloVe: Global
vectors for word representation. In Empiri-
cal Methods in Natural Language Processing
(EMNLP), pages 1532–1543. https://doi
.org/10.3115/v1/D14-1162
Agnieszka
Pietraszkiewicz,
Magdalena
Formanowicz, Marie Gustafsson Send´en,
Ryan L. Boyd, Sverker Sikstr¨om, and Sabine
Sczesny. 2019. The big two dictionaries:
Capturing agency and communion in natural
language. European Journal of Social Psychol-
ogy, 49(5):871–887. https://doi.org
/10.1002/ejsp.2561
Lisa Posch, Arnim Bleier, Fabian Fl¨ock, and
Markus Strohmaier. 2018. Characterizing the
global crowd workforce: A cross-country com-
parison of crowdworker demographics.
In
Eleventh International AAAI Conference on
Web and Social Media.
Vinodkumar Prabhakaran, Ben Hutchinson, and
Margaret Mitchell. 2019. Perturbation sensitiv-
ity analysis to detect unintended model biases.
In Proceedings of the 2019 Conference on Em-
pirical Methods in Natural Language Process-
ing and the 9th International Joint Conference
on Natural Language Processing (EMNLP-
IJCNLP), pages 5740–5745, Hong Kong,
China. Association for Computational Lin-
guistics. https://doi.org/10.18653
/v1/D19-1578
Vinodkumar Prabhakaran, Aida Mostafazadeh
Davani, and Mark Diaz. 2021. On releasing
annotator-level labels and information in data-
sets. In Proceedings of The Joint 15th Lin-
guistic Annotation Workshop (LAW) and 3rd
Designing Meaning Representations (DMR)
Workshop, pages 133–138, Punta Cana,
Dominican Republic. Association for Compu-
tational Linguistics.
Ashwin Rajadesingan, Reza Zafarani, and Huan
Liu. 2015. Sarcasm detection on twitter: A
behavioral modeling approach. In Proceedings
of the eighth ACM international conference on
web search and data mining, pages 97–106.
https://doi.org/10.1145/2684822
.2685316
Georg Rasch. 1993. Probabilistic Models for
Some Intelligence and Attainment Tests. ERIC.
Bj¨orn Ross, Michael Rist, Guillermo Carbonell,
Benjamin Cabrera, Nils Kurowsky,
and
Michael Wojatzki. 2017. Measuring the reli-
ability of hate speech annotations: The case of
the european refugee crisis. In Proceedings of
the Workshop on Natural Language Processing
for Computer Mediated Communication.
Paul Rottger, Bertie Vidgen, Dirk Hovy, and
Janet Pierrehumbert. 2022. Two contrasting
317
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
the North American Chapter of
data annotation paradigms for subjective NLP
tasks. In Proceedings of the 2022 Conference
of
the As-
sociation for Computational Linguistics: Hu-
man Language Technologies, pages 175–190,
Seattle, United States. Association for Compu-
tational Linguistics. https://doi.org/10
.18653/v1/2022.naacl-main.13
Maarten Sap, Dallas Card, Saadia Gabriel, Yejin
Choi, and Noah A. Smith. 2019. The riskofracial
bias in hate speech detection. In Proceedings of
the 57th Annual Meeting of the Association for
Computational Linguistics, pages 1668–1678.
Maarten Sap, Swabha Swayamdipta, Laura
Vianna, Xuhui Zhou, Yejin Choi, and Noah
Smith. 2022. Annotators with attitudes: How
annotator beliefs and identities bias toxic lan-
guage detection. In Proceedings of the 2022
Conference of the North American Chapter
of the Association for Computational Linguis-
tics: Human Language Technologies, Seattle,
United States. Association for Computational
Linguistics.
Steven E. Stemler and Adam Naples. 2021. Rasch
measurement v. item response theory: Knowing
when to cross the line. Practical Assessment,
Research & Evaluation, 26:11.
Nathaniel Swinger, Maria De-Arteaga, Neil
Thomas Heffernan IV, Mark DM Leiserson,
and Adam Tauman Kalai. 2019. What are the bi-
ases in my word embedding? In Proceedings of
the 2019 AAAI/ACM Conference on AI, Ethics,
and Society, pages 305–311. https://doi
.org/10.1145/3306618.3314270
Zeerak Talat. 2016. Are you a racist or am I see-
ing things? Annotator influence on hate speech
detection on twitter. In Proceedings of the first
workshop on NLP and computational social
science, pages 138–142.
Zeerak Talat, Smarika Lulz, Joachim Bingel, and
Isabelle Augenstein. 2021. Disembodied ma-
chine learning: On the illusion of objectivity in
NLP. Anonymous preprint under review.
Alexandra N. Uma, Tommaso Fornaciari, Dirk
Hovy, Silviu Paun, Barbara Plank, and Massimo
Poesio. 2021. Learning from disagreement: A
survey. Journal of Artificial Intelligence Re-
search, 72:1385–1470. https://doi.org
/10.1613/jair.1.12752
Ameya Vaidya, Feng Mai, and Yue Ning. 2020.
Empirical analysis of multi-task learning for re-
ducing identity bias in toxic comment detection.
In Proceedings of the International AAAI Con-
ference on Web and Social Media, volume 14,
pages 683–693.
Bertie Vidgen, Tristan Thrush, Zeerak Talat,
and Douwe Kiela. 2021. Learning from the
worst: Dynamically generated datasets to im-
prove online hate detection. In Proceedings of
the 59th Annual Meeting of the Association
for Computational Linguistics and the 11th
International Joint Conference on Natural Lan-
guage Processing (Volume 1: Long Papers),
pages 1667–1682, Online. Association for
Computational Linguistics. https://doi.org
/10.18653/v1/2021.acl-long.132
Claudia Wagner, Markus Strohmaier, Alexandra
Olteanu, Emre Kıcıman, Noshir Contractor, and
Tina Eliassi-Rad. 2021. Measuring algorithmically
infused societies. Nature, 595(7866):197–204.
https://doi.org/10.1038/s41586
-021-03666-1, PubMed: 34194046
Maximilian Wich, Hala Al Kuwatly, and Georg
Groh. 2020. Investigating annotator bias with
a graph-based approach. In Proceedings of the
fourth workshop on online abuse and harms,
pages 191–199.
Thomas Wolf, Lysandre Debut, Victor Sanh,
Julien Chaumond, Clement Delangue, Anthony
Moi, Pierric Cistac, Tim Rault, Remi Louf,
Morgan Funtowicz, Joe Davison, Sam Shleifer,
Patrick von Platen, Clara Ma, Yacine Jernite,
Julien Plu, Canwen Xu, Teven Le Scao, Sylvain
Gugger, Mariama Drame, Quentin Lhoest,
and Alexander Rush. 2020. Transformers:
State-of-the-art natural language processing. In
Proceedings of the 2020 Conference on Em-
pirical Methods in Natural Language Pro-
cessing: System Demonstrations, pages 38–45,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653/v1
/2020.emnlp-demos.6
Mengzhou Xia, Anjalie Field,
and Yulia
in
Tsvetkov. 2020. Demoting racial bias
In Proceedings of
hate speech detection.
the Eighth International Workshop on Natu-
ral Language Processing for Social Media,
318
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
pages 7–14, Online. Association for Compu-
tational Linguistics.
Huhhot, China. Chinese Information Process-
ing Society of China.
Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun.
2021. A robustly optimized BERT pre-training
approach with post-training. In Proceedings
of the 20th Chinese National Conference on
Computational Linguistics, pages 1218–1227,
Linda X. Zou and Sapna Cheryan. 2017. Two axes
of subordination: A new model of racial posi-
tion. Journal of Personality and Social Psychol-
ogy, 112(5):696–717. https://doi.org/10
.1037/pspa0000080, PubMed: 28240941
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
5
0
2
0
7
5
7
3
0
/
/
t
l
a
c
_
a
_
0
0
5
5
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
319
