ArticlePDF Available

How Language Shapes Prejudice Against Women: An Examination Across 45 World Languages

Authors:

Abstract and Figures

Language provides an ever-present context for our cognitions and has the ability to shape them. Languages across the world can be gendered (language in which the form of noun, verb, or pronoun is presented as female or male) versus genderless. In an ongoing debate, one stream of research suggests that gendered languages are more likely to display gender prejudice than genderless languages. However, another stream of research suggests that language does not have the ability to shape gender prejudice. In this research, we contribute to the debate by using a Natural Language Processing (NLP) method which captures the meaning of a word from the context in which it occurs. Using text data from Wikipedia and the Common Crawl project (which contains text from billions of publicly facing websites) across 45 world languages, covering the majority of the world’s population, we test for gender prejudice in gendered and genderless languages. We find that gender prejudice occurs more in gendered rather than genderless languages. Moreover, we examine whether genderedness of language influences the stereotypic dimensions of warmth and competence utilizing the same NLP method.
Content may be subject to copyright.
1
1
How Language Shapes Prejudice Against Women: An Examination
Across 45 World Languages
David DeFranza
1
Himanshu Mishra
2
Arul Mishra
3
Forthcoming: Journal of Personality and Social Psychology (Attitudes and Social Cognition)
1
University of Utah, david.defranza@eccles.utah.edu
2
University of Utah, himanshu.mishra@utah.edu
3
University of Utah, arul.mishra@utah.edu
2
2
Abstract
Language provides an ever-present context for our cognitions and has the ability to shape them.
Languages across the world can be gendered (language in which the form of noun, verb, or
pronoun is presented as female or male) versus genderless. In an ongoing debate, one stream of
research suggests that gendered languages are more likely to display gender prejudice than
genderless languages. However, another stream of research suggests that language does not have
the ability to shape gender prejudice. In this research, we contribute to the debate by using a
Natural Language Processing (NLP) method which captures the meaning of a word from the
context in which it occurs. Using text data from Wikipedia and the Common Crawl project
(which contains text from billions of publicly facing websites) across 45 world languages,
covering the majority of the world’s population, we test for gender prejudice in gendered and
genderless languages. We find that gender prejudice occurs more in gendered rather than
genderless languages. Moreover, we examine whether genderedness of language influences the
stereotypic dimensions of warmth and competence utilizing the same NLP method.
Keywords: gender, prejudice, stereotype, natural language processing, word embeddings, text
analysis
3
3
Language is the ever-present context of our cognitions and provides an insight into our
cognition (i.e., thoughts, attitudes, and beliefs; Carnaghi & Maass, 2007; Collins & Clément
2012). It is considered both a vessel for holding cognitions as well as a lens that can direct (or
distort) those cognitions (Sutton 2010). Differences in how concepts and objects are defined
across languages, therefore, leads to differences in how they are perceived by respective speakers
(Casasanto 2008). One dimension where languages differ is gender. World languages can be
categorized as gendered (a language in which the form of noun, verb, or pronoun is presented as
female or male) versus genderless languages. This raises the following question: if gender is
emphasized continuously in language, then is it possible that gendered languages hold differing
perceptions of women relative to men compared to genderless languages.
There are conflicting answers to this question. One stream of work shows that people
who speak gendered languages do display greater prejudice against women compared to those
who speak genderless languages (Wasserman & Weseley, 2009). This stream of research
suggests that genderedness of language can cause perpetuation of gender perceptions and aligns
with findings on linguistic relativity (Casasanto, 2008, 2016; Oh 2003). However, a competing
argument is that language cannot influence perceptions against women because language merely
reflects the beliefs of a culture and cannot influence or cause gender prejudice (Deutscher, 2010;
Fodor, 1975; Pinker, 1994).
We contribute to this debate in three ways. First, we compare 45 world languages
covering the majority of the world’s nations, to test for gender prejudice across several gendered
and genderless languages. Second, instead of testing for prejudice through attitudinal measures
or through response time tests, we utilized thousands of gigabytes of data, voluntarily generated
4
4
by people across different languages and different contexts, when they may not be specifically
discussing gender issues. Wikipedia and a corpus of web crawl data from over five billion web
pages, known as the Common Crawl, serve as our data source.
Consistent with past work, we define prejudice as the affective component of intergroup
bias where social groups are differentially associated with positively and negatively valanced
words (Fiske, Cuddy, Glick, & Xu, 2002; Bhatia, 2017; Todd, Simpson, Thiem, & Neel, 2016).
Such valenced associations have been shown to automatically affect subsequent behavior when
certain social categories are activated (Allport, 1954; Devine, 1989). Knowing the nature of the
diverging association is important (e.g. greater association of positive with women or with men)
because only then successful interventions can be designed. Therefore, third, we investigate
whether gender prejudice occurs because of a greater prejudice against women or a lesser
prejudice for men.
We next present the details of the debate as to whether language can shape
cognitions or whether it can only communicate cognitions, and then discuss how we attempt to
resolve this debate.
Theoretical Review
The Debate: Does Language Shape or Simply Communicate Thought?
Linguistic relativity. The Sapir-Whorf hypothesis claims that the language one
speaks influences that individual’s perception of the world. When people speak different
languages, they think differently (Casasanto, 2008; Kay & Kempton, 1984). In its strongest
formulation, the Sapir-Whorf hypothesis is a type of linguistic determinism, which suggests that
the categories made available by a person’s language determines that person’s thoughts (Kay &
5
5
Kempton, 1984; Pinker, 1994). Therefore, it is assumed that language, thought, and culture are
intertwined and each language creates a distinctive worldview (Gumperz & Levinson, 1991). A
more common and modern version of the Sapir-Whorf hypothesis is known as linguistic
relativity (Casasanto, 2008, 2016; Pinker, 1994). It states that “differences among languages
cause differences in the thoughts of their speakers” (Pinker, 1994, p. 57) because language
provides an ever-present context for each of our cognitions and hence has the ability to shape
cognitions (Casasanto, 2008, 2016). In other words, instead of focusing on language and its
influence on vague notions of worldview, research on linguistic relativity now focuses on
investigating the influence of language on specific cognitive effects like perception of time and
motion (Casasanto, 2008).
For example, research has demonstrated that English motion verbs focus on the
manner of motion (rolling, climbing, running) while Spanish motion verbs focus on the path of
motion (enter versus exit). Such a difference in encoding results in individuals remembering
more manner-related versus path-related information depending on which language they use
(Casasanto, 2016; Oh, 2003).
One challenge to such research is that it must disentangle cognition from language.
In order to ensure that the task used to evaluate cognitions is not related to language, studies have
used an extralinguistic feature. The extralinguistic feature is orthogonal to language (i.e., it does
not have a language tag attached to it) but is still affected depending on which language an
individual speaks. Hence, it provides converging, empirical evidence for the ability of language
to shape cognition (Casasanto, 2008). For instance, researchers used a combined temporal-spatial
assessment task in which participants saw a line being drawn between two points (spatial) over a
6
6
period of time (temporal). Participants were then asked to reproduce either the temporal or the
spatial aspect of the line. English-speaking participants, for whom time was related to length
(e.g., a long versus short time), confounded spatial and temporal tasks. However, Greek-speaking
participants, for whom time is associated with volume (e.g., big versus small), made fewer
mistakes and were able to disentangle temporal and spatial tasks (Casasanto, 2008, 2016). When
the spatial-temporal task was converted into an assessment of volume with time, Greek speakers
made more mistakes than English speakers. Therefore, the use of an extralinguistic feature
provides support for linguistic relativity. A similar stream of literature examines the role
gendered versus genderless languages on gender prejudice.
Gendered versus genderless languages and prejudice against women. Grammatical
gender describes the formal rules of a language such as the form of nouns, verbs, or pronouns
that reflect gender and distinguish world languages into three broad categories: gendered
languages, natural gendered languages, and genderless languages (Prewitt-Freilino, Caswell, &
Laakso, 2012; Stahlberg, Braun, Irmen, & Sczesny, 2007). In gendered languages, rules of
grammar dictate that most nouns appear with a male or female gender (e.g., la pelota or “ball” in
Spanish which has a female designation or le ballon in French, which has a male designation).
Their associated pronouns, adjectives, and verbs reflect the assigned gender. Examples of such
languages include Spanish, French, Russian, and Hindi, to name only a few. Generally, in natural
gender languages, nouns tend not to have gender tags but gender is indicated through pronouns
such as she, he, or it. English is perhaps the best example of a natural gender language.1
Genderless languages do not have any grammatical rules to distinguish nouns by gender and
7
7
their pronouns relating to men and women could be the same (Stahlberg et al., 2007). Examples
of genderless languages include Chinese and Finnish.
Gendered languages personify even inanimate objects with a specific male or female
gender, and such personification could unconsciously reinforce attitudes about gender
(Boroditsky & Schmidt, 2000; Jakobson, 1959). For instance, in languages where the sun is
given a male gender (e.g., French and Spanish), it is perceived to be more powerful and
courageous but when it is given a female gender (e.g., German), it is considered warm and
nurturing (Hofstätter, 1963). Assignment of gender to objects brings the gender marker into the
foreground and results in greater salience of gender roles. It follows that such salience would
result in perceptions of that gender being reiterated more frequently. Thus, in Spanish and French
the moon, which has a female gender assignment, is seen as being weaker and more passive than
the sun because of the salience of the perception of women being weaker (Boroditsky &
Schmidt, 2000; Hofstätter, 1963).
Gendered (as opposed to genderless) languages influence not only the perception of
women but also attitudes about them, which can result in gender prejudice (Boroditsky &
Schmidt, 2000). For instance, utilizing EEG, it has been demonstrated that languages can
influence cognitions non-consciously, since one uses language continuously and the influence
can occur without awareness (Boutonnet, Athanasopoulos, & Thierry, 2012). In countries of the
world where gendered language is spoken, women earn lower wages (van der Velde, Tyrowicz,
& Siwinska, 2015) and face more barriers to participating in the political arena (Santacreu-Vasut,
Shoham, & Gay, 2013).
8
8
Language merely reflects thought. A competing argument, which forms the other
part of the debate, is that language cannot influence the perpetuation of gender prejudice because
it can communicate thought but it cannot shape cognitions (Cela-Conde & Marty, 1998; Pinker,
1994). For instance, researchers from this side of the argument suggest that any cognition can be
expressed in any language (Cela-Conde & Marty, 1998; Hunt & Agnoli, 1991; Pinker, 1994).
English speakers, for example, may not have a dedicated word for schadenfreude but it is easy
enough to communicate the experience of relishing in another’s misfortune (Deutscher, 2010).
Similarly, we can see all languages adjust and adapt to describe new ideas, objects, and
innovations (Hunt & Banaji, 1988; Hunt & Agnoli, 1991). Hence, they argue that Whorf’s strict
relativism, which claims language dictates what its speakers are capable of thinking or
understanding cannot be true. This stream of research also argues that it takes far less time to
coin and adopt a term that defines a new concept but far longer for the concept to form and
acquire an entity of its own (Hunt & Banaji, 1988; Hunt & Agnoli, 1991; January & Kako, 2007;
McWhorter, 2016). That is, language provides a label to a construct but does not shape the
construct (Deutscher, 2010).
However, it is worth noting that many of these arguments get reduced to strict
pronouncements that only language can shape cognition or that language can only reflect
cognition. Some research has argued that a middle ground may be more reflective of reality
wherein language is influenced by changing cognition of people but in return language can also
affect how people perceive and especially express thoughts (Boroditsky & Schmidt, 2000;
Casasanto, 2008; Deutscher, 2010; Kay & Kempton, 1984; Sera, Elieff, Forbes, Burch,
Rodríguez, & Dubois, 2002). While language may not constrain cognition, it does dictate the
9
9
means through which that cognition can be expressed. For instance, if there is no best-fitting
word to express a cognition in a language, then the repeated use of the only available word for
that cognition could result in the thought becoming modified because of the meaning that word
conveys. German speakers, for example, are forced to describe a fork as die gabel, which is
female (Deutscher, 2010). While the gender marker may not have literal meaning to German
speakers, there is evidence that this subtle gender cue leads to associations of a fork with female
characteristics (Boutonnet et al., 2012; Hofstätter, 1963; Sera et al., 2002).
In sum, there is an ongoing debate whether language shapes cognitions (linguistic
relativity) or merely reflects cognitions. Research has used gendered versus genderless language
to test for gender prejudice. This stream of research has demonstrated the greater presence of
gender prejudice in gendered compared to genderless language as evidence of language
influencing cognition (Boyd & Richerson, 1985; Evans & Levinson, 2009; Konishi, 1993; Liu &
Karahanna, 2017; Sera, Berge, & del Castillo-Pintado, 1994; Stubbs, 1996). Such a pattern of
results could be because socially assigned gender roles of men (e.g., being stronger) versus
women (e.g., being weaker) are emphasized more often in gendered languages. However,
research examining the association between gender prejudice and the use of gendered versus
genderless language has been criticized on a few fronts.
One criticism is that the evidence rests on studies considering only two or three
languages. Since there are more than 100 different languages spoken by 85% of the world
population (Simmons & Fennig, 2018), the conclusions drawn from two or three languages
might not be generalizable. It is possible that observed effects might disappear when other or
more languages are considered. A second criticism is that such research often uses explicit
10
10
measures, which can allow people to suppress their perceptions or they may not be aware that
they hold a certain perception (Banaji, Hardin, & Rothman, 1993; Greenwald & Banaji, 1995;
Greenwald et al., 2002; Greenwald, McGhee, & Schwartz, 1998; Greenwald, Nosek, & Banaji,
2003). Finally, there is very little research that describes the nature of gender prejudice in
gendered languages, which makes it difficult to conclude whether the prejudice is predominately
against women or in favor of men.
We contribute to this debate by studying 45 world languages in order to examine if
gendered languages are more likely to contain gender prejudice. This helps address the criticism
of using only one or two languages while testing. We used a natural language processing (NLP)
method to test whether genderedness of language is associated with gender prejudice.
If past research showing the relationship of language genderedness and gender prejudice
is to hold true, we should find gender prejudice to be more prevalent in gendered rather than
genderless languages. Conversely, if it is true that language does not influence gender prejudice,
but is simply a reflection of the culture of that language community (Cralley & Ruscher, 2005;
Stahlberg et al., 2007), then we should find gender prejudice across both gendered and
genderless languages because gender prejudice is known to exist in every part of the world
(Matthes, Prieler, & Adam, 2016; Prewitt-Freilino et al., 2012; United Nations Development
Programme, 2017; Williams, Satterwhie, & Best, 1999; World Economic Forum, 2017).
Existing studies examining gender prejudice in gendered versus genderless languages
have directly queried participants’ regarding their gender beliefs. For instance, questions such as
whether men and women should be treated similarly or paid the same for the same job or
whether women have already been given too much by the government tend to elicit responses
11
11
that capture explicit gender prejudice (Wasserman & Weseley, 2009). But some research has
suggested that direct questioning cannot reliably capture prejudice or latent traits that inform
behavior (Banaji et al., 1993; Deutscher, 2010; Sterling, Jost, & Bonneau 2020). This could
happen because individuals tend to respond in a socially desirable manner where they hide
gender prejudice (Zerbe & Paulhus, 1987). Secondly, they may engage in self-deception and not
believe that they hold any gender prejudice. Finally, sampling from a small set of languages and
querying participants directly may lack ecological validity necessary to test psychological
constructs (Sterling et al., 2020).
We used an NLP method that takes care of shortcomings of prior investigations. It
allowed us to use the voluntarily produced thoughts and feelings of people to examine the
association of language genderedness and gender prejudice. The voluntarily produced text comes
from large corpora, Wikipedia and Common Crawl. These text corpora reflect the real-world
communications and opinions produced by people online. They are a compilation of text
generated by individuals on a vast range of topics, and not just when they are specifically
discussing gender issues. Briefly, the intuition behind our approach is the following. If a
language contains prejudice against women, then words representing women compared to those
representing men will appear semantically more associated with negatively valanced words.
The challenge is that it is not always easy to assess semantic association among
words. To this end, the NLP method we used can analyze vast quantities of text (in the context of
this manuscript it is several thousand gigabytes of data) to capture semantic association among
words. Moreover, it is important to note that while text corpora allow us to measure associations
between gender and prejudice-exhibiting words, these associations do not suggest a causal link
12
12
between genderedness of language and prejudice. However, past research across domains, has
demonstrated that cooccurrence studies showing a link between two variables in the real world
can be used in deriving many useful insights; we do the same in this research (Maki, 1979; Maki,
Maki, & Marsh, 1977).
We next explain this method and how it is able to capture semantic associations
from the context of the word and convert it into a numeric format that can be used in further
analysis.
Method
In today’s world, with people producing vast amounts of digital text, collections of
such text, or corpora, have become the holding body of people’s cognitions, feelings, and actions
across time and regions. Analysis of digitized text has provided insights in the social (e.g.,
examining the condition of marginalized groups; Ben-David & Matamoros-Fernández, 2016;
Garg, Schiebing, Jurafsky, & Zou, 2018; Thelwall, Wilkinson, & Uppal, 2010), business (e.g.,
analyzing consumer reviews to extract quality inference; Tirunillai & Tellis, 2014), political (e.g.,
analyzing speeches, legislative addresses, or parliamentary debates to examine influence of laws;
Miller, 2013; Schonhardt-Bailey, 2006; Stewart & Zhukov, 2009), and legal (e.g., approval of
home loans or using license plates to predict crime; Schrage, 2014) domains. In our research, we
use a specific NLP method that creates word embeddings and thus enables study of semantic
associations among words.
What are Word Embeddings?
We can easily discern that the colors red and maroon are more similar than the
colors red and yellow. However, when trying to discern if “rose” is more similar to “pleasant” or
13
13
“unpleasant,” we will hesitate a bit because without a context it is difficult to understand if
“rose” is a flower or an individual. This poses a fundamental question: How do words acquire
meaning? Saussure (1916/1959) argued that “language is a system of interdependent terms in
which the value of each term results solely from the simultaneous presence of others.
Wittgenstein (1953) asserted that “the meaning of words lies in their use.” Inspired by these
assertions linguists developed the distributional hypothesis (Harris, 1954) captured in Firth’s
(1957) oft-quoted statement, “you shall know a word by the company it keeps.” That is, words
derive their meaning from the context in which they occur. For instance, the meaning of “spring”
as a season or a piece of coiled metal can only be ascertained by the context in which it is used.
Similarly, pianos are heavy in the context of moving furniture, but they are musical in the context
of a virtuoso.
The distributional hypothesis proposes a link between how words are distributed in
language and the similarity in their meaning (Jurafsky & Martin, 2019). Consider reading the
following sentences: hikareti is bad for your lungs; hikareti is addictive; hikareti contains
nicotine. Even if one is not familiar with Maori language, one can infer that hikareti means
something that is used to inhale smoke like a cigarette or cigar. One can infer the meaning of
hikareti from words that appeared in its context because these words also appear in the context of
cigarette or cigar. Therefore, by examining different words (nicotine, addictive, lungs) that
appear in the context of a word (hikareti) one can estimate whether it is similar or dissimilar to
other words (cigarette vs. apple).
The essence of the distributional hypothesis is that words, which share similar context
words are likely to be semantically similar. Note, in this work semantic similarity between words
14
14
implies relatedness in their meaning (i.e., likeness of their meaning) because they occur with
similar context words, similarity does not imply sharing similar features. When we have large
quantities of text, the goal is to let algorithms use the distributional hypothesis to learn semantic
associations among words like a human can learn about hikareti and cigarettes/cigar. Modern
NLP algorithms do this by building word embedding models where each word in a large text
corpus is represented as a multidimensional vector (commonly referred to as word embeddings
or word vectors). That is, word embeddings are the numeric, multidimensional representation of
words.
By learning relationships based on the context in which words occur, word embeddings
allow complex comparisons, like estimating the semantic similarity between words, to be
performed using relatively simple computations. For instance, the cosine of the angle between
embeddings of two words informs us how similar they are (Bolukbasi, Chang, Zou, Saligrama, &
Kalai, 2016; Caliskan, Bryson, & Narayanan, 2017). If represents the embedding of word
and represents the embedding of word then similarity (S) between and is:
(Equation 1)
In Equation 1, the numerator represents the dot product or inner product of the
embeddings and , while the denominator is the product of their respective Euclidean
lengths.
This similarity calculation offers some interesting insights. When two embeddings are
parallel, they have an angle of zero degrees and a cosine of one; they are considered maximally
similar in meaning. When two word embeddings are perpendicular, thus having an angle of 90
1w
v
1
w
2w
v
2
w
1
w
2
w
12
12
12
()
||||
,ww
ww
v
S
v
v
w
v
w=!
1w
v
2w
v
15
15
degrees and a cosine of zero, they are considered orthogonal (unrelated) in meaning. When two
word embeddings have an angle of 180 degrees and a cosine of -1, they are considered opposite
in meaning. Beyond similarity, word embeddings also preserve linguistic regularity that humans
use in answering analogies like “King:Men::Women:__”. Thus, embeddings of words can also be
used to answer analogy tasks.
Embeddings of words that preserve their semantic relationships align with our goal of
examining gender prejudice in gendered versus genderless languages. If a language is
documenting prejudice against women then words representing women compared to those
representing men will appear semantically more related with negatively valanced words.
Therefore, we utilized embeddings created with one of the latest machine learning methods,
fastText, which accommodates morphological variations across different languages. fastText
(Bojanowski, Grave, Joulin, & Mikolov, 2016; Grave, Bojanowski, Gupta, Joulin, & Mikolov,
2018; Mikalov, Grave, Bojanowski, Puhrsch, & Joulin, 2017) modifies the Word2Vec algorithm
used in prior work (Mikolov, Chen, Corrado, & Dean, 2013; Mikolov, Sutskever, Chen, Corrado,
& Dean, 2013).
The intuition behind the method is the following. If we define the context as certain
number of words occurring before or after a target word, then the goal is to predict a target word
from the words that occur in its context. As an example, take the sentence “farmers grow Arabica
coffee beans in Colombia.” If the context window is three then fastText uses {Farmers, grow,
Arabica, beans, in, Colombia} to predict {coffee}. If we denote target word coffee as and the
set of context words as set then probability that is the target word given will be
. However, for an irrelevant target word n (e.g., glacier, also called a negative sample),
t
w
t
C
t
w
t
C
16
16
will not provide a good prediction and the probability of predicting an irrelevant target like n
denoted by will be lower than predicting .
Extending this example, we see that if two words occur in the same context (e.g.,
Colombia and coffee) then their similarity will be greater than two words that do not appear in
the same context (e.g., Colombia and glacier). Therefore, will be a function of the
similarity between and . Similarly, will be a function of similarity between
and n. For a relevant target word , the algorithm also samples irrelevant words, like n, and
maximizes the similarity between the context and relevant target word while minimizing the
similarity between the context and sampled irrelevant words.
Formally, we can describe the above algorithm in the following manner. Given words
, where is a set of context words and is the target word (e.g., coffee),
the similarity (S) will be . Here ‘. is dot product, is the embedding of word
, and is average of the embeddings for each word in . The probability that is a target
word given context words is a function of similarity S. The function that turns similarity to
probability is the sigmoid or logistic function (Jurafsky & Martin, 2019). Therefore,
(Equation 2)
Which is equivalent to:
(Equation 3)
t
C
(| )
t
Pn C
t
w
t
C
(| )
t
Pn C
t
C
t
w
T
123
{, , , ,}
t
ww w w
t
C
t
w
(, )
t T wt ct
Sw C v v=!
wt
v
t
w
ct
v
t
C
t
w
t
C
(,)
1
1
(|) tt
tt SwC
pC e
w-
=
+
(,)
1
log ( ) log 1
|tt
tt SwC
pw Ce-
=
+
17
17
Similarly, when word n is drawn from a negative sample (e.g., glacier) then the probability of n
occurring in the context window is:
(Equation 4)
and the probability of n not being a target word in the context of is where,
(Equation 5)
If is the set of negatively sampled words and if we assume these negatively
sampled words to be independent, then the probability of all the words in not being a target
word in the context of is obtained by multiplying the individual probability of each n:
(Equation 6)
Now we have two probabilities, one is the probability that is a target word in the
context of and the other is the probability that negatively sampled words n are not target
words in the context of . Both of these probabilities are functions of similarity. Thus, the
(, )
1
(| ) 1t
tSnC
pC e
n-
=
+
t
C
1(|)
t
pn C-
(, )
1
1(|)1
1t
tSnC
pn C e-
-=-
+
(, )
(, )
1
t
t
SnC
SnC
e
e
-
-
+
=
(, )
1
1t
SnC
e
=
+
c
N
c
N
t
C
(, )
1( |) 1
1t
ct SnC
N
e
Pn C-=Î
+
Õ
(, )
1
log(1 ( | )) log 1
c
t
tSnC
nN
c
pn N C e
Î
=
+
å
t
w
t
C
t
C
18
18
learning objective becomes maximizing both of these probabilities. For all the words T, this
objective can be written as maximizing:
(Equation 7)
This maximization process is governed by a shallow two-layer neural network. Starting
with a hidden layer of randomly initialized embeddings, the text corpus is streamed into the
network one context window at a time. Embeddings are updated after each instance with the goal
of maximizing similarity between context words and relevant target words while minimizing
similarity between context words and negative sample words (Bojanowski et al., 2016; Grave et
al., 2018).
However, the above approach of creating word embeddings has a limitation: it ignores
the internal structure of words. By creating an embedding for each word, the algorithm does not
consider that in morphologically rich languages, various forms of a word can occur. For instance,
in French or Spanish, most verbs have more than forty different inflected forms, while the
Finnish language has fifteen cases for nouns (Bojanowski et al., 2016). In Russian, there are at
least 70 word forms for dog, encoding gender, age, number, sentiment and other semantic
connotations (Vylomova, Cohn, He, & Haffari, 2016). Many of these forms rarely occur in the
text corpus, thus reducing the quality of the resulting word embeddings (i.e., their ability to
capture semantic relationship).
fastText addresses this weakness by utilizing subword information to capture common
and systematic morphological rules (Bojanowski et al., 2016; Grave et al., 2018; Joulin, Grave,
(,) (,)
1
11
log log
11
tt t
c
T
SwC SnC
tnN
ee
-
=Î
éù
+
êú
++
ëû
åå
19
19
Bojanowski, & Mikolov, 2016). The word, “where,” for example, will be represented in
word2vec by just one set of d-dimensional embeddings. However, in fastText, “where” will be
decomposed into various subgroups of n contiguous characters known as character n-grams
(Bojanowski et al., 2016; Grave et al., 2018; Joulin et al., 2016). When n is set to three, as is
common, “where” will be decomposed into subgroups of characters { wh, whe, her, ere, re }.
Each of these subgroups will have their own d-dimensional embeddings and the final embedding
of the complete word “where” will be averaged to represent the individual word. Formally, a
word is represented by:
(Equation 8)
This new representation of essentially changes the similarity calculation, which now becomes
. As an added benefit, fastText also improves the performance of
embeddings when the corpus contains misspelled words or concatenations of words (Bojanowski
et al., 2016; Grave et al., 2018; Joulin et al., 2016).
The dimensions of the word embedding actually do not have any meaning on their
own. In other words, they don’t represent any specific quality or attribute of a word. These
embeddings are numerical values (vectors) assigned to each word that preserve their semantic
relationship with other words. Three hundred dimensions of word embeddings, instead of 500 or
200 dimensions, are used since it provides the optimal performance in finding semantic
associations (Bojanowski et al., 2016; Grave et al., 2018).
t
w
1
wt n
nN
xv
NÎ
+å
t
w
1
(, )( )
t t wt n ct
nN
Sw C v x v
NÎ
=+
å
20
20
Identifying gender prejudice. To determine whether the fastText embeddings are
detecting gender prejudice from text corpora, we adapted a method suggested by Caliskan, et al.
(2017). Since, fastText converts a word into a high dimension numerical vector, we can then find
the level of similarity between two words using cosine similarity, as described in Equation 1. The
resulting score captures the similarity between two words (i.e., whether a target word is more or
less similar to a valenced word).
Consider, for example, how word embeddings can be used to test for gender
prejudice. Let us assume that there are two sets of male (e.g., he, him, man) and female (e.g., she,
her, woman) words. These gender words are compared to two sets of valenced words: one
positive (e.g., freedom, health, love, peace) and the other negative (e.g., abuse, crash, filth,
sickness). If a language contains prejudice against women then male words would display more
similarity to positive rather than negative words compared to female words. Formally, we can
represent the presence of a gender prejudice as a net similarity measure where S represents
similarity (as in Equation 1).
S(male, positive) – S(male, negative) > S(female, positive) – S(female, negative) (Equation 9)
The left-hand side of the inequality in Equation 9 measures the net similarity of
male words to positively versus negatively valenced words. The term on the right-hand side of
the inequality measures the net similarity of female words to positive versus negative words.
Estimating these net similarities helps us both detect whether gender prejudice exists and what is
the nature of the gender prejudice. To detect gender prejudice, we compare the left-hand side of
the inequality with the right-hand side. If the left-hand side is more than the right-hand side then
it would indicate that the net association of male words with positive words is more than the net
21
21
association of female words with positive words. Another way of representing this is by taking
the difference of the right-hand-side from the left-hand side; higher value of the difference would
indicate lower prejudice against men. That is, it shows that males are more positively represented
in text corpora than females. Such a discrepancy would be surprising in light of the Pollyanna
hypothesis (Boucher & Osgood, 1969) also known a linguistic positivity bias (Iliev, Hoover,
Dehghani, & Axelrod, 2016). The Pollynana hypothesis states that positive words are more
prevalent, easily learned, and used across languages. People tend to couch communications in
positive language, even when the core content of the communication is negative in nature.
Studies of large text corpora, using translations across languages, have supported this hypothesis
(Dodds et al., 2015). Thus, the finding that despite their prevalence, positive words are
semantically more associated with males than females provides evidence of prejudice.
Further, Equation 9 can help in understanding the nature of the prejudice. For instance,
the left-hand side of the inequality could be more than the right-hand side if a) male words were
roughly equal in similarity to positive and negative words but female words were much more
similar to negative words than positive words; b) male words were more similar to positive than
negative words but female words were much more similar to negative words than positive words;
or c) male words were more similar to positive than negative words but female words were
roughly equal in similarity to positive and negative words. Following Equation 9 we can answer
what kind of association between gender words and valenced words leads to prejudice.
Details of the measurement of similarity and the statistical tests are provided in Appendix
B.
Data Description
22
22
We used two text corpora to test if gender prejudice is more likely to exist in
gendered languages. The first corpus consisted of data from the free online encyclopedia
Wikipedia (Bojanowski et al., 2016) and the second was from the Common Crawl project, which
contains snapshots of all the text that can be scraped from the publicly facing Internet, since
2013 (Grave et al., 2018) and contains more than 630 billion words. These two sources are
available in several world languages and tend to cover a broad set of topics. Moreover,
Wikipedia is considered nominally objective and fact-based (Filipacchi, 2013; Glott, Schmidt, &
Ghosh, 2010; Gruwell, 2015; Hill & Shaw, 2013; Pfeil, Zaphiris, & Ang, 2006; Shane-Simpson
& Gillespie-Lynch, 2017; Wagner, Garcia, Jadidi, & Strohmaier, 2015; Wagner, Graells-Garrido,
Garcia, & Menczer, 2016) and thus provides a conservative setting to test for gender prejudice.
For both data sets, we used pre-trained word embeddings in our analysis. The
Wikipedia2 and Common Crawl3 embeddings were produced by Bojanowski et al. (2016) and
Grave et al. (2018) respectively and are publicly available. Each embedding consisted of a vector
for each word, derived from the fastText algorithm (Bojanowski et al., 2016; Grave et al., 2018;
Joulin et al., 2016) discussed previously. In both data sets, the word embedding for each word
consisted of 300 dimensions.
Gender and Valence Words
For our examination of gender prejudice, we used two categories of words: gender
words and valence words. Gender words consisted of 218 noun and pronoun pairs identified and
tested by Bolukbasi et al. (2016). For instance: he, his, him, man, men versus she, hers, her,
woman, women; refer to Appendix C for the complete list. Valence words consisted of 25
positively and 25 negatively valenced words commonly used to assess prejudice (Caliskan et al.,
23
23
2017; Greenwald et al., 1998, 2003). The relative similarities (see Equation 1), between the
gender words and the valenced words were used to measure the strength of association (see
Equation 9) and hence, the level of gender prejudice. Details of the lexicon are available in
Appendix C.
Language Selection and Validation
We considered several criteria in order to select languages for analysis. First, since
one of the embeddings was based on Wikipedia corpora, languages were selected based on their
popularity on Wikipedia (Wikipedia, 2018a). Due to the differential properties and behaviors of
various Wikipedia communities, there is a trade off in many languages between volume of
articles and the depth of the articles. Volume describes the total number of unique articles in the
project. Depth is a metric created by Wikipedia to compare the frequency of updates across a
project, normalizing for the size and popularity of different projects (Wikipedia, 2018b). We
selected languages based on the volume of available data, rather than focusing on those
languages that have the most intensely updated and edited articles because we were primarily
interested in the natural occurrence of gender prejudice in language. Using these criteria as a
guide, the initial search produced 60 languages with 100,000 or more Wikipedia articles.
Since gender words and valence words used in prior work are in English, we
translated them using the Google Translate API using a procedure outlined in prior work (Dodds
et al., 2015). While the coverage of Google Translate service is significant, it is not
comprehensive. As a result, our list of 60 languages was reduced to 49 to match the coverage of
Google Translate. Despite this reduction, the 49 languages spanned various geographic regions
and language families. Translation was conducted on every gender and valence word for each
24
24
language, as well as on a validation lexicon. The translated words were used to extract vectors
from the appropriate language embedding.
Next, a validation method was implemented to test how good the word embeddings
were in detecting semantic similarity in the corpora we were using in the analysis. First, a
validation word list was constructed consisting of generic positive nouns (e.g., beauty,
celebration, smile) and generic negative nouns (e.g., drought, emergency, fatality). The selected
words were unambiguous and based on similar validation lexicons used in past research
(Caliskan et al., 2017; Nosek, Banaji, & Greenwald, 2002; Nosek et al., 2009), but adapted to
avoid associations with specific cultural constructs. The validation word list was translated, both
Wikipedia and Common Crawl embeddings were obtained for each translated word, and
associations were calculated using cosine similarity.
We hypothesized that a test of our method would require known positive words to
have a statistically significant net positive similarity (i.e., known positive words should be more
similar to positively valenced words), while known negative words would have a statistically
significant net negative similarity (i.e., known negative words should be more similar to
negatively valenced words). For example, the association between “hero” and “love” was
expected to be greater than the association between “hero” and “filth” (see Appendix C for a full
list of the validation words). If the hypothesized relationship (i.e., known positive words are
more similar to positively valenced words and known negative words are more similar to
negatively valenced words) was not statistically significant, the language was considered to have
failed validation and was removed.
25
25
Five languages, Cebuano, Japanese, Latin, Vietnamese, and Welsh failed validation
and were removed from the Wikipedia data set and four languages, Cebuano, Hebrew, Latin, and
Vietnamese, failed validation and were removed from the Common Crawl data set.4 Despite the
reduction due to translation availability and validation, languages common on every continent,
including first or second languages of more than 50% of the world population were included
(Simmons & Fennig, 2018). As discussed previously and suggested by past research (Corbett,
1991; Stahlberg et al., 2007), we divided languages based on whether they are gendered or
genderless. To assign languages to these groups, we first consulted the World Atlas of Language
Structures (Haspelmath, 2005), which provided the necessary information for more than half of
the languages in our data set. Next, we consulted Ethnologue: Languages of the World (Simmons
& Fennig, 2018), which accounted for the majority of remaining languages. Finally, any
remaining unclassified languages were validated through references in the series Gender Across
Languages (Hellinger & Bußmann, 2001). After translation, validation, and classification, there
were 28 languages in the gendered group and 16 languages in the genderless group for the
Wikipedia data set and 28 languages in the gendered group and 17 languages in the genderless
group for the Common Crawl data set. Across both data sets, we considered 45 unique languages
(see Appendix A for more details on the validation and classification procedures).5 Moreover, we
also controlled for various geographic, demographic, and cultural covariates in our analysis (see
Appendix E).
Results
Gender Prejudice
26
26
We tested for gender prejudice in the two corpora, Wikipedia and Common Crawl,
where we compared the similarity of target male words (e.g., he, him) and female words (e.g.,
she, her) with positive (e.g., love, pleasure, lucky) as well as negatively valenced words (e.g.,
abuse, sickness, ugly).
Wikipedia. The analysis revealed a statistically significant gender prejudice in the
Wikipedia data set for 21 of the 44 validated languages, χ2(1) = 3.8724, p = 0.0491. Specifically
the analysis revealed that 17 gendered languages (60%) exhibited gender prejudice while only
four (25%) genderless languages did so.
We al so examined whether net gender prejudice for males or females interacted with
gendered versus genderless languages. Recall from Equation 9, gender prejudice for male is
S(male, positive) – S(male, negative) and for female is S(female, positive) – S(female, negative).
As mentioned earlier, a higher value of the difference of the right-hand-side from the left-hand
side would indicate lower prejudice against men. A one-way MANOVA with 2 (type of
prejudice: male vs. female) dependent variables x 2 (genderedness of languages: gendered vs.
genderless) independent variables demonstrated a significant interaction F(2, 41) = 6.1939, p =
0.0045. On decomposing this interaction, we find that in genderless languages, there is no
difference in net prejudice for male words (M = 0.007) or for female words (M = 0.0048; t =
0.6057, p > 0.05). However, among gendered languages, the net prejudice was lower against
male words (M = 0.0111) than against female words (M = 0.0048; t = 0.6057, p < 0.001).
Therefore, gender prejudice occurs in gendered language because of the greater association of
male words (compared to female words) with positively valenced words (see Figure 1).
27
27
Figure 1. Nature of gender prejudice in the Wikipedia data set.
Please refer to Table S2 in Appendix D where similarity estimates (S(male, positive),
S(male, negative), S(female, positive), and S(female, negative)) across all languages are reported.
Common Crawl
We conducted the same analysis using Common Crawl word embeddings. As with
the Wikipedia data set, we find that gendered languages exhibited significant gender prejudice
compared to genderless languages (see Figure 2). 19 gendered languages exhibited gender
prejudice (67%), however, none of the genderless languages exhibited gender prejudice, χ2(1) =
17.281, p 0.
28
28
Similar to the Wikipedia analysis, the interaction of net prejudice against male or
female words with gendered versus genderless languages was significant F(2, 42) = 9.212, p <
0.001. On decomposing this interaction, we find that in genderless languages, there is no
difference in net prejudice against male words (M = 0.0165) or against female words (M =
0.0145; t = 0.4744, p > 0.05). However, among gendered languages the net prejudice was lower
against male words (M = 0.0229) than against female words (M = 0.0072; t = 5.504, p < 0.001).
This pattern is consistent with the Wikipedia analysis and again indicates that male words
(compared to female words) had a higher net similarity to positive words in gendered but not in
genderless languages (see Figure 2). Therefore, our results demonstrate that not only is gender
prejudice more prevalent in gendered languages, it is caused by a higher association of male
words with positive words (see Figure 3 and Figure 4).
29
29
Figure 2. Nature of the gender prejudice in the Common Crawl data set.
30
30
Figure 3. Gender prejudice in languages in Wikipedia corpora
31
31
Figure 4. Gender prejudice in languages in Common Crawl corpora
Additional tests
It is worth noting that languages can belong to different language families (e.g.
Indo-European, Sino-Tibetan). Therefore, one could argue that the difference we are observing
between gendered and genderless languages is not because of genderedness but because of the
32
32
language family. To address this account, we conducted an analysis using a Hierarchical Linear
model (HLM) that controlled for variance in each language family, as well as across. The results
indicated that genderedness still had a significant influence on net gender prejudice in both the
Wikipedia and Common Crawl data (detailed analysis are reported in Appendix E).
One could also argue that there may be specific features of the region where
languages are spoken that could affect gender prejudice over and above genderedness of
languages. In order to address such a concern we ran our analysis controlling for geographic,
cultural, and demographic covariates that have been shown to influence gender equality. The
results indicate that for both Wikipedia and Common Crawl corpora genderedness significantly
predicted gender prejudice for each language (detailed analysis are reported in Appendix E).
Our examination so far focused on prejudice, which is the affective component of
the intergroup bias. Our findings provide some insights into the debate as to whether gendered
languages display prejudice against women. However, there is another component of the
intergroup bias, the cognitive component, known as stereotyping (Fiske et al., 2002; Bhatia,
2017; Todd et al., 2016). Stereotypes have been defined as specific beliefs about a group, such as
descriptions of what members of a particular group look like, how they behave, or their abilities.
They include cognitive representations of how members of a group are similar to one another
and different from members of other groups. Hence, a gender stereotype would be an association
of male or female with different semantic traits (e.g., home or work, art or science, warmth or
competence).
We next examine whether genderedness of languages can affect the associations of
men versus women with the traits of warmth and competence.
33
33
Gender Stereotypes
Past research has categorized gender stereotypes along two general dimensions:
warmth and competence (Abele et al., 2016; Durante et al., 2013; Durante et. al. 2017; Fiske,
Cuddy, & Glick, 2007; Rosenberg, Nelson, & Vivekananthan, 1968; Rudman, Greenwald, &
McGhee, 2001). Although both are positive in nature, research has considered whether women
might be stereotypically more associated with warmth compared to men while men are believed
to be more associated with competence. Moreover, it has been demonstrated that the warmth-
competence dimensions occur across cultures and regions (Durante et. al., 2017). Given the
broad presence of the warmth-competence dimensions, we examined whether women were
associated more with warmth and men more with competence. And importantly whether such
stereotypic associations were stronger in gendered languages compared to genderless languages.
Similar to the gender prejudice analysis, this analysis utilized a lexicon of words to
represent gender constructs (male and female), validated in previous research (Bolukbasi et al.,
2016). Since, no single lexicon has been developed to clearly represent the warmth-competence
constructs, we drew from established descriptions from Rosenberg et al. (1968), Rudman et al.
(2001), Fiske et al. (2007), and Abele et al. (2016), to build a representative construct dictionary
for warmth (e.g. sincere, warm, caring) and competence (e.g. capable, competent, skillful).
Details of this lexicon can be found in Appendix C.
We tested for gender stereotypes in the two corpora, Wikipedia and Common Crawl,
by comparing the similarity of target male and female words (e.g., he, she, her, him) to warmth
words (e.g. friendly, warm, good natured) as well as competence words (e.g. skillful, confident,
competent). It is important to note that when we analyzed gender prejudice, we had positively
34
34
and negatively valenced words. Hence, we were able to calculate a net similarity as described in
Equation 9. Since competence and warmth are both positively valenced, creating a net difference
score like S(male, warmth) – S(male, competence) or S(female, warmth) – S(female,
competence) would not be meaningful. Therefore, we analyzed the competence and warmth
dimensions separately.
Common Crawl.
For the construct of competence, a MANOVA revealed a significant interaction
between male or female words with gendered versus genderless languages, F(2, 42) = 13.077, p
< 0.001). Decomposing the interaction across gendered and genderless languages, we find that in
genderless languages there is no difference in similarity or association of competence words with
male words (M = 0.219) or female words (M = 0.211; t = -0.1997, p > 0.05). However, among
gendered languages, there is a greater association of competence words with male words (M =
0.182) than female words (M = 0.157; t = -4.6437, p < 0.001) showing that men are associated
more with competence than women in gendered languages (see Figure 5).
35
35
Figure 5. Association of male and female words with the construct of competence in gendered
versus genderless languages in Common Crawl.
Similarly for the construct of warmth in the Common Crawl data, a MANOVA
revealed a significant interaction between male or female words with gendered versus genderless
languages, F(2, 42) = 13.599, p < 0.001. On decomposing the interaction, we find that in
genderless languages, there is no difference in association with warmth words for male words (M
= 0.233) or for female words (M = 0.234; t = 0.0445, p > 0.05). However, among gendered
languages, male words (M = 0.194) are more associated with warmth words than female words
(M = 0.182; t = -2.1421, p = 0.0367; see Figure 6). This means that male words are associated
more with warmth than female words in gendered languages.
36
36
Figure 6. Association of warmth with male versus female words in Common Crawl.
Across both warmth and competence dimensions we see an interesting pattern.
Males are more associated with both competence and warmth in gendered languages, which
could be because both are positively valenced.
Wikipedia.
For the construct of competence, results were similar to those from Common Crawl.
A MANOVA revealed a significant interaction F(2, 41) = 3.8329, p = 0.0298. On
decomposition, we find that there is no difference in association with competence words for male
words (M = 0.238) or for female words (M = 0.229; t = -0.1527, p > 0.05) in genderless
languages. However, among gendered languages, male words were more associated with
37
37
competence words (M = 0.201) than female words (M = 0.184; t = -2.2439, p = 0.0291; see
Figure 7).
Figure 7. Association of competence with male and female words in Wikipedia data in gendered
versus genderless languages.
For the construct of warmth, a MANOVA revealed a significant interaction F(2, 41)
= 5.1579, p = 0.01. Specifically on decomposing the interaction, we find that there is no
difference in association with warmth for male words (M = 0.254) or for female words (M =
0.258; t = 0.0545, p > 0.05) in genderless languages. Among gendered languages, there was no
significant difference in association with warmth with male words (M = 0.217) or for female
words (M = 0.213; t = -0.4966, p > 0.05; see Figure 8).
38
38
Figure 8. Association of male and female words with the construct of warmth in gendered versus
genderless languages in Wikipedia.
Similar to the analysis we conducted with gender prejudice, we examine the
influence of gendered versus genderless languages by controlling for language families in a
HLM model. As in the previous analysis of prejudice, we see a significant influence of
genderedness when the influence of individual language families are considered. We present the
details of the analysis in Appendix F.
Discussion
Our findings contribute to the debate of whether the genderedness of language is
associated with gender prejudice. One stream of research suggests that the genderedness of
39
39
language is associated with gender prejudice while another stream of research suggests that
language merely reflects the beliefs of a culture and cannot be associated with gender prejudice.
If the latter argument were to hold true, we should have seen the prevalence of gender prejudice
equally across languages. Instead, in our examination of large text corpora of 45 world
languages, we find that gender prejudice appears more in gendered rather than genderless
languages.
Our findings indicate that when gender is made more salient through the
genderedness of language, it results in certain associations of that gender being reiterated more
frequently. Unfortunately, we find that associations of men tend to be more positive (i.e., there is
a closer association of male words with positive words) than women. Moreover, past research
has examined gender prejudice in English (Caliskan et al., 2017) and across time (Garg et al.,
2018), suggesting that languages can effectively be used to test for gender prejudice. However, to
date, this approach has not been used comparatively across a large group of common world
languages. Using languages that span continents and cultures, we find evidence that gender
prejudice exists more in gendered languages.
Second, instead of testing for gender prejudice through self-reports or direct
questioning, we tested for gender prejudice using millions of pages of text data. We used
Wikipedia and Common Crawl corpora that contains text (e.g., facts, thoughts, opinions)
voluntarily produced by millions of people. By measuring the association of words voluntarily
produced across 45 different world languages when people were not specifically being queried
about gender prejudice, gives us a glimpse into how prevalent gender prejudice is. We
demonstrate that gender prejudice exists cross-culturally. We also show, in line with past
40
40
research, albeit using more languages and using text data, that prejudice present in language has
the potential to influence future thoughts. Wikipedia and Common Crawl are not just repositories
of people’s thought but are used by them to form future opinions and hence, can result in
perpetuation of gender prejudice in gendered languages.
Third, knowing the nature of the prejudice can help in developing successful
interventions. Therefore, it is important to know whether gender prejudice exists in certain
languages but it is equally important to know whether the prejudice occurs because of a greater
prejudice against women or a lesser prejudice against men. Our results across both data sets
indicate that the gender prejudice is driven more by the greater association of positive words
with men, which would result in their being evaluated favorably, in many significant decisions.
For instance, two equally qualified individuals, one female and one male, may be treated
differently just because of their gender. As mentioned earlier, such a finding is troubling in light
of the “Pollyanna” effect (Boucher & Osgood, 1969; Dodd et al., 2015), in which individuals
have a higher tendency to use positive words, because it shows that the association of positive
words with women is weaker. Figure 3 and Figure 4 demonstrate that the positive-negative
association of men and women varies significantly across languages. Understanding the nature of
this association will be critical for future efforts and policies focused on eliminating gender
prejudice around the world.
Limitations and Future Research
In our research we tested for the presence of gender stereotypes using the
dimensions of warmth and competence. Our findings indicate that men are associated more with
both warmth and competence than women in gendered languages, despite the fact that past
41
41
research has shown that women are stereotypically associated with warmth. Moreover, we do not
find stereotypic associations in genderless languages. Our conjecture is that such a pattern of
results could occur because of the following. It is possible that what we are observing in the text
corpora is distinct from commonly studied psychological stereotyping and is instead a specific
construct that is encoded linguistically by which males are consistently associated with positive
attributes. That is, although women are typically reported to be more closely associated with
warmth when such responses are directly elicited from participants, language usage shows a
different relationship. Language usage might be denigrating women such that positively valenced
words are more strongly associated with men than women. In this scenario, our analysis is
revealing a linguistic pattern of association which is stronger in gendered languages. Future
research could examine the association of men with positive constructs, generally, across
different languages and cultures.
We examined only warmth and competence but other stereotypic dimensions such as
weakness versus power or home versus career that have been stereotypically associated with one
gender can also be examined. Future research can also consider other dimensions such as
perceptions of emotional versus rational decision-making and test whether gendered versus
genderless languages show a differing pattern of association.
It is also worth noting that we find evidence consistent with predictions made in
prior work on linguistic relativity. However, since we are using associations to infer relationship,
our evidence does not directly test for the tenet of linguistic relativity: that language can affect
cognitions. Future research can consider other extra-linguistic perception to cast additional light
on the language-shapes-cognition versus language-reflects-cognition debate. The findings of this
42
42
research focus on written text. Given the prevalence of spoken along with written words, future
research should consider whether genderedness of language affects prejudice or stereotype in
spoken language.
On a broader level, social psychology research can use the method employed in this
manuscript in various ways. First, the large text corpora of Wikipedia and Common Crawl can be
used to replicate many existing findings in social psychology. As an example, one could examine
whether promotion-focused language does indeed occur more with aspirational words while
prevention-focused language occurs more with responsibility words. This would help establish
robustness of existing findings in an ecologically valid data set in which people were not directly
responding to a question but were instead providing their opinion on a diverse set of topics.
Second, other text corpora exist that range across time such as news articles or twitter feeds.
These can be used to gain novel insights into how events occurring across time are influencing
different psychological constructs. For example, such a method could be used to chart the rise of
populist discourse and associated changes in the perception of minorities or women. Finally,
word embedding methods represent just one type of text analysis tool capable of modeling
semantic associations. Hence, word embeddings are best suited to testing for the association
between psychological constructs. However, simpler text analysis methods such as topic
modelling (e.g., Latent Dirichlet Allocation; LDA) that rely on word frequency can also be very
useful to examine what topics are most likely in a document. For instance, LDA analysis of right
and left leaning media outlets might reveal a difference in topics popular on such venues.
Subsequent word embedding analysis could offer insight into the association these topics might
have with specific individuals, events, or groups. In this way, word embedding analysis becomes
43
43
an important part of an ensemble of methods that together provide deeper insight into, greater
generalizability of, and a more robust understanding of psychological phenomena.
44
44
References
Abele, A. E., Hauke, N., Peters, K., Louvet, E., Szymkow, A., & Duan, Y. (2016). Facets of the
fundamental content dimensions: Agency with competence and assertiveness—
communion with warmth and morality. Frontiers in Psychology, 7, 1-17.
Allport, G. W. (1954). The nature of prejudice. Cambridge, MA: Addison-Wesley Publishing
Company.
Banaji, M. R., Hardin, C., & Rothman, A. J. (1993). Implicit stereotyping in person judgment.
Journal of Personality and Social Psychology, 65(2), 272–281.
Ben-David, A., & Matamoros-Fernández, A. (2016). Hate speech and covert discrimination on
social media: Monitoring the Facebook pages of extreme-right political parties in Spain.
International Journal of Communication, 10, 1167–1193.
Bhatia, S. (2017). The semantic representation of prejudice and stereotypes. Cognition, 164, 46-
60.
Bojanowski, P., Grave, E., Joulin, A., & Mikolov, T. (2016). Enriching word vectors with
subword information. ArXiv:1607.04606 [Cs]. Retrieved from
http://arxiv.org/abs/1607.04606
Bolukbasi, T., Chang, K.-W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer
programmer as woman is to homemaker? Debiasing word embeddings. In D. D. Lee, M.
Sugiyama, U. V. Luxburg, I. Guyon, & R. Garnett (Eds.), Advances in Neural
Information Processing Systems (Vol. 29, pp. 4349–4357). Retrieved from
http://papers.nips.cc/paper/6228-man-is-to-computer-programmer-as-woman-is-to-
homemaker-debiasing-word-embeddings.pdf
Boroditsky, L., & Schmidt, L. A. (2000). Sex, syntax, and semantics. Proceedings of the Annual
Meeting of the Cognitive Science Society, 22, 16.
45
45
Boucher, J., & Osgood, C. E. (1969). The Pollyanna hypothesis. Journal of Verbal Leaning and
Verbal Behavior, 8(1), 1–8.
Boutonnet, B., Athanasopoulos, P., & Thierry, G. (2012). Unconscious effects of grammatical
gender during object categorisation. Brain Research, 1479, 72–79.
Boyd, R., & Richerson, P. J. (1985). Culture and the Evolutionary Process. Chicago, IL:
University of Chicago Press.
Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from
language corpora contain human-like biases. Science, 356(6334), 183–186.
Carnaghi, A., & Maass, A. (2007). In-group and out-group perspectives in the use of derogatory
group labels: Gay versus fag. Journal of Language and Social Psychology, 26(2), 142-
156.
Casasanto, D. (2008). Who’s Afraid of the Big Bad Whorf? Crosslinguistic differences in
temporal language and thought. Language Learning, 58(s1), 63–79.
Casasanto, D. (2016). Linguistic relativity. In N. Riemer (Ed.), The Routledge Handbook of
Semantics (pp. 158–174). New York, NY: Routledge.
Cela-Conde, C. J., & Marty, G. (1998). Noam Chomsky’s minimalist program and the
philosophy of mind. An interview. Syntax, 1(1), 19–36.
Collins, K. A., & Clément, R. (2012). Language and prejudice: Direct and moderated effects.
Journal of Language and Social Psychology, 31(4), 376-396.
Corbett, G. G. (1991). Gender. New York: Cambridge University Press.
Cralley, E. L., & Ruscher, J. B. (2005). Lady, girl, female, or woman: Sexism and cognitive
busyness predict use of gender-biased nouns. Journal of Language and Social
Psychology, 24(3), 300–314.
Deutscher, G. (2010). Through the language glass: Why the world looks different in other
languages. New York, NY: Metropolitan Books.
46
46
Devine, P. G. (1989). Stereotypes and prejudice: Their automatic and controlled components.
Journal of Personality and Social Psychology, 56(1), 5–18.
Dodds, P. S., Clark, E. M., Desu, S., Frank, M. R., Reagan, A. J., Williams, J. R.,…Danforth, C.
M. (2015). Human language reveals a universal positivity bias. Proceedings of the
National Academy of Sciences, 112(8), 2389-2394.
Durante, F., Fiske, S. T., Gelfand, M. J., Crippa, F., Suttora, C., Stillwell, A.,…Teymoori, A.
(2017). Ambivalent stereotypes link to peace, conflict, and inequality across 38 nations.
Proceedings of the National Academy of Sciences, 114(4), 669-674.
Durante, F., Fiske, S. T., Kervyn, N., Cuddy, A. J. C., Akande, A., Adetoun, B. E. (2013).
Nations’ income inequality predicts ambivalence in stereotype content: How societies
mind the gap. British Journal of Psychology, 52(4), 726-746.
Evans, N., & Levinson, S. C. (2009). The myth of language universals: Language diversity and
its importance for cognitive science. Behavioral and Brain Sciences, 32(05), 429.
Filipacchi, A. (2013, April 24). Wikipedia’s sexism toward female novelists. The New York
Times. Retrieved from https://www.nytimes.com/2013/04/28/opinion/sunday/wikipedias-
sexism-toward-female-novelists.html
Firth, J. R. (1957). A synopsis of linguistic theory, 1930-1955. Studies in Linguistic Analysis,
1957, 1–32.
Fiske, S. T., Cuddy, A. J. C., & Glick, P. (2007). Universal dimensions of social cognition:
Warmth and competence. Trends in Cognitive Sciences, 11(2), 77–83.
Fiske, S. T., Cuddy, A. J. C., Glick, P., & Xu, J. (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), 878-902.
Fodor, J. A. (1975). The Language of Thought. New York, NY: Crowell.
47
47
Garg, N., Schiebinger, L., Jurafsky, D., & Zou, J. (2018). Word embeddings quantify 100 years
of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences,
115(16), E3635–E3644.
Glott, R., Schmidt, P., & Ghosh, R. (2010). Wikipedia survey: Overview of results (pp. 1–11).
Retrieved from
http://www.ris.org/uploadi/editor/1305050082Wikipedia_Overview_15March2010-
FINAL.pdf
Grave, E., Bojanowski, P., Gupta, P., Joulin, A., & Mikolov, T. (2018). Learning word vectors for
157 languages. ArXiv:1802.06893 [Cs]. Retrieved from http://arxiv.org/abs/1802.06893
Greenwald, A. G., & Banaji, M. R. (1995). Implicit social cognition: Attitudes, self-esteem, and
stereotypes. Psychological Review, 102(1), 4–27.
Greenwald, A. G., Banaji, M. R., Rudman, L. A., Farnham, S. D., Nosek, B. A., & Mellott, D. S.
(2002). A unified theory of implicit attitudes, stereotypes, self-esteem, and self-concept.
Psychological Review, 109(1), 3–25.
Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. K. (1998). Measuring individual differences
in implicit cognition: The implicit association test. Journal of Personality and Social
Psychology, 74(6), 1464–1480.
Greenwald, A. G., Nosek, B. A., & Banaji, M. R. (2003). Understanding and using the Implicit
Association Test: I An improved scoring algorithm. Journal of Personality and Social
Psychology, 85(2), 197–216.
Gruwell, L. (2015). Wikipedia’s politics of exclusion: Gender, epistemology, and feminist
rhetorical (in)action. Computers and Composition, 37, 117–131.
Gumperz, J. J., & Levinson, S. C. (1991). Rethinking linguistic relativity. Current Anthropology,
32(5), 613–623.
Harris, Z. S. (1954). Distributional structure. Word, 10(2–3), 146–162.
48
48
Haspelmath, M. (2005). World Atlas of Language Structures. Oxford, UK: Oxford University
Press.
Hellinger, M. (2001). English: Gender in a global language. In M. Hellinger & H. Bußmann
(Eds.), Gender across languages: The linguistic representation of women and men (pp.
105-113). Amsterdam: J Benjamins Publishing Company.
Hellinger, M., & Bußmann, H. (Eds.). (2001). Gender across languages: The linguistic
representation of women and men. Philadelphia, PA: J. Benjamins Publishing Company.
Hill, B. M., & Shaw, A. (2013). The Wikipedia gender gap revisited: Characterizing survey
response bias with propensity score estimation. PLOS ONE, 8(6), e65782.
Hofstätter, P. R. (1963). Über sprachliche Bestimmungsleistungen: Das Problem des
grammatikalischen Geschlechts von Sonne und Mond. [On linguistic performances of
connotation: The problem of the grammatical gender of sun and moon.]. Zeitschrift Für
Experimentelle Und Angewandte Psychologie, 10(1), 91–108.
Hunt, E., & Banaji, M. R. (1988). The Whorfian hypothesis revisited: A cognitive science view
of linguistic and cultural effects on thought. In J. W. Berry, S. H. Irvine, & E. B. Hunt
(Eds.), Indigenous Cognition: Functioning in Cultural Context (pp. 57–84). Dordrecht,
Netherlands: Martinus Nijhoff Publishers.
Hunt, Earl, & Agnoli, F. (1991). The Whorfian hypothesis: A cognitive psychology perspective.
Psychological Review, 98(3), 377–389.
Iliev, R., Hoover, J., Dehghani, M., & Axelrod, R. (2016). Linguistic positivity in historical texts
reflects dynamic environmental and psychological factors. Proceedings of the National
Academy of Sciences, 113(49), E7871–E7879.
Jakobson, R. (1959). On the linguistic aspects of translation. In R. A. Brower (Ed.), On
Translation (pp. 232–239). Cambridge, MA: Harvard University Press.
49
49
January, D., & Kako, E. (2007). Re-evaluating evidence for linguistic relativity: Reply to
Boroditsky (2001). Cognition, 104(2), 417–426.
Joulin, A., Grave, E., Bojanowski, P., & Mikolov, T. (2016). Bag of tricks for efficient text
classification. ArXiv:1607.01759 [Cs]. Retrieved from http://arxiv.org/abs/1607.01759
Jurafsky, D., & Martin, J. H. (2019). Speech and language processing (3rd ed.). Retrieved from
https://web.stanford.edu/~jurafsky/slp3/
Kay, P., & Kempton, W. (1984). What is the Sapir-Whorf hypothesis? American Anthropologist,
New Series, 86(1), 65–79.
Konishi, T. (1993). The semantics of grammatical gender: A cross-cultural study. Journal of
Psycholinguistic Research, 22(5), 519–534.
Liu, Q. B., & Karahanna, E. (2017). The dark side of reviews: The swaying effects of online
product reviews on attribute preference construction. MIS Quarterly, 41(2), 427-A6.
Maki, W. S. (1979). Discrimination learning without short-term memory: Dissociation of
memory processes in pigeons. Science, 204(4388), 83-85.
Maki, R. H., Maki, W. S., & Marsh, L. G. (1977). Generalization of above/below, right/left
effects to compass directions and natural stimuli. Bulletin of the Psychonometric Society,
10(4), 307-310.
Matthes, J., Prieler, M., & Adam, K. (2016). Gender-role portrayals in television advertising
across the globe. Sex Roles, 75(7–8), 314–327.
McWhorter, J. H. (2016). The language hoax: Why the world looks the same in any language.
New York, NY: Oxford University Press.
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word
representations in vector space. ArXiv:1301.3781 [Cs]. Retrieved from
http://arxiv.org/abs/1301.3781
50
50
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed
representations of words and phrases and their compositionality. Advances in Neural
Information Processing Systems, 26(NIPS 2013), 9.
Mikolov, T., Grave, E., Bojanowski, P., Puhrsch, C., Joulin, A. (2017). Advances in pre-training
distributed word representations. ArXiv:1712.09405 [Cs]. Retrieved from
https://arxiv.org/abs/1712.09405
Miller, M. C. (2013). Wronged by empire: Post-imperial ideology and foreign policy in India
and China. Stanford, CA: Stanford University Press.
Nosek, B. A., Banaji, M. R., & Greenwald, A. G. (2002). Harvesting implicit group attitudes and
beliefs from a demonstration web site. Group Dynamics: Theory, Research, and Practice,
6(1), 101–115.
Nosek, B. A., Smyth, F. L., Sriram, N., Lindner, N. M., Devos, T., Ayala, A., … Kesebir, S.
(2009). National differences in gender: Science stereotypes predict national sex
differences in science and math achievement. Proceedings of the National Academy of
Sciences, 106(26), 10593-10597.
Oh, K. J. (2003). Language, cognition, and development: Motion events in English and Korean.
University of California, Berkeley.
Pew Research Center. (2014). Religious diversity index scores by country. Retrieved from
http://www.pewforum.org/2014/04/04/religious-diversity-index-scores-by-country/
Pfeil, U., Zaphiris, P., & Ang, C. S. (2006). Cultural differences in collaborative authoring of
Wikipedia. Journal of Computer-Mediated Communication, 12(1), 88–113.
Pinker, S. (1994). The Language Instinct. New York: W. Morrow and Co.
Prewitt-Freilino, J. L., Caswell, T. A., & Laakso, E. K. (2012). The gendering of language: A
comparison of gender equality in countries with gendered, natural gender, and genderless
languages. Sex Roles, 66(3–4), 268–281.
51
51
Rosenberg, S., Nelson, C., & Vivekananthan, P. S. (1968). A multidimensional approach to the
structure of personality impressions. Journal of Personality and Social Psychology, 9(4),
283–294.
Rudman, L. A., Greenwald, A. G., & McGhee, D. E. (2001). Implicit self-concept and evaluative
implicit gender stereotypes: Self and ingroup share desirable traits. Personality and
Social Psychology Bulletin, 27(9), 1164–1178.
Santacreu-Vasut, E., Shoham, A., & Gay, V. (2013). Do female/male distinctions in language
matter? Evidence from gender political quotas. Applied Economics Letters, 20(5), 495–
498.
Saussure, F. de (1959). Course in general linguistics. (W. Baskin, Trans.). New York, NY:
McGraw-Hill. (Original work published 1916).
Schonhardt-Bailey, C. (2006). From the corn laws to free trade: Interests, ideas, and institutions
in historical perspective. Cambridge, MA: MIT Press.
Schrage, M. (2014, January 29). Big data’s dangerous new era of discrimination. Retrieved from
https://hbr.org/2014/01/big-datas-dangerous-new-era-of-discrimination
Sera, M. D., Berge, C. A. H., & del Castillo-Pintado, J. (1994). Grammatical and conceptual
forces in the attribution of gender by English and Spanish speakers. Cognitive
Development, 9(3), 261–292.
Sera, M. D., Elieff, C., Forbes, J., Burch, M. C., Rodríguez, W., & Dubois, D. P. (2002). When
language affects cognition and when it does not: An analysis of grammatical gender and
classification. Journal of Experimental Psychology: General, 131(3), 377–397.
Shane-Simpson, C., & Gillespie-Lynch, K. (2017). Examining potential mechanisms underlying
the Wikipedia gender gap through a collaborative editing task. Computers in Human
Behavior, 66, 312–328.
52
52
Simmons, G. F., & Fennig, C. D. (2018). Ethnologue: Languages of the World (21st ed.).
Retrieved from https://www.ethnologue.com/
Stahlberg, D., Braun, F., Irmen, L., & Sczesny, S. (2007). Representation of the sexes in
language. In K. Fiedler (Ed.), Social Communication (pp. 163–187). New York, NY:
Psychology Press.
Sterling, J., Jost, J. T., & Bonneau, R. (2020). Political psycholinguistics: A comprehensive
analysis of the language habits of liberal and conservative social media users. Journal of
Personality and Social Psychology.
Stewart, B. M., & Zhukov, Y. M. (2009). Use of force and civil–military relations in Russia: An
automated content analysis. Small Wars & Insurgencies, 20(2), 319–343.
Stubbs, M. (1996). Text and corpus analysis: Computer-assisted studies of language and culture.
Oxford, OX, UK; Cambridge, MA: Blackwell Publishers.
Sutton, R. M. (2010). The creative power of language in social cognition and intergroup
relations. In G. Howard, S. Reid, & J. Hardwood (Eds.), Dynamics of intergroup
communication (pp. 105-115). New York, NY: Peter Lang.
Tajick, S. (2018). STC climate index. Retrieved from http://globalresidenceindex.com/hnwi-
index/climate-index/
Thelwall, M., Wilkinson, D., & Uppal, S. (2010). Data mining emotion in social network
communication: Gender differences in MySpace. Journal of the American Society for
Information Science and Technology, 61(1), 190–199.
Tirunillai, S., & Tellis, G. J. (2014). Mining marketing meaning from online chatter: Strategic
brand analysis of big data using latent dirichlet allocation. Journal of Marketing
Research, 51(4), 463–479.
53
53
Todd, A. R., Simpson, A. J., Thiem, K. C., & Neel, R. (2016). The generalization of implicit
racial bias to young black boys: Automatic stereotyping or automatic prejudice? Social
Cognition, 34(4), 306-323.
Transparency International. (2016). Corruption perceptions index. Retrieved from
https://www.transparency.org/news/feature/corruption_perceptions_index_2016
United Nations Development Programme. (2017). Human Development Report 2016: Human
Development for Everyone.
van der Velde, L., Tyrowicz, J., & Siwinska, J. (2015). Language and (the estimates of) the
gender wage gap. Economics Letters, 136, 165–170.
Vylomova, E., Cohn, T., He, X., & Haffari, G. (2016). Word representation models for
morphologically rich languages in neural machine translation. ArXiv:1606.04217 [Cs].
Retrieved from http://arxiv.org/abs/1606.04217
Wagner, C., Garcia, D., Jadidi, M., & Strohmaier, M. (2015). It’s a man’s Wikipedia? Assessing
gender inequality in an online encyclopedia. ArXiv:1501.06307 [Cs]. Retrieved from
http://arxiv.org/abs/1501.06307
Wagner, C., Graells-Garrido, E., Garcia, D., & Menczer, F. (2016). Women through the glass
ceiling: gender asymmetries in Wikipedia. EPJ Data Science, 5(1), 5.
Wasserman, B. D., & Weseley, A. J. (2009). ¿Qué? Quoi? Do languages with grammatical gender
promote sexist attitudes? Sex Roles, 61(9–10), 634–643.
Wikimedia Foundation. (2018). Wikimedia traffic analysis report: Page views per Wikipedia
language. Retrieved from https://stats.wikimedia.org/archive/squid_reports/2018-
05/SquidReportPageViewsPerLanguageBreakdown.htm
Wikipedia. (2018a). List of Wikipedias. In Wikipedia. Retrieved from
https://en.wikipedia.org/w/index.php?title=List_of_Wikipedias&oldid=860318950
54
54
Wikipedia. (2018b). Wikipedia article depth. Retrieved from
https://meta.wikimedia.org/wiki/Wikipedia_article_depth
Williams, J. E., Satterwhite, R. C., & Best, D. L. (1999). Pancultural gender stereotypes
revisited: The five factor model. Sex Roles, 40(7–8), 513–525.
Wittgenstein, L. (1973) Philosophical investigations (3rd ed.; G.E.M. Anscombe, Trans.).
Oxford, UK: Pearson.
World Economic Forum. (2017). The Global Gender Gap Report: 2017. Geneva: World
Economic Forum.
Zerbe, W. J., & Paulhus, D. L. (1987). Socially desirable responding in organizational behavior:
A reconception. Academy of Management Review, 12(2), 250–264.
55
Footnotes
1 For the purposes of this research, we classify languages based on the assignment of
gender markers to nouns, verbs, and adjectives. As such, English is classified as a genderless
language, which is consistent with past literature that utilizes a binary taxonomy (Hellinger,
2001).
2 https://github.com/facebookresearch/fastText/blob/master/docs/pretrained-
vectors.md (Wikipedia)
3 https://github.com/facebookresearch/fastText/blob/master/docs/crawl-
vectors.md (Common Crawl)
4 While it is difficult to assess exactly why these languages failed validation, it could
be due to failed or imprecise translations. Future work could address this challenge by
developing specific lexicons for these languages which are drawn from, translated, and validated
by native speakers of the language.
5 A power analysis confirmed that the resulting group sizes would be sufficient to
identify medium-sized effects, if such effects existed.
56
SUPPLEMENTAL MATERIAL
57
Appendix A: Validation
A validation method was implemented to verify the ability of the word embedding
algorithm to correctly identify semantic associations in each language. A validation lexicon was
constructed consisting of generic positive nouns (e.g., beauty, fun, hero) and generic negative
nouns (e.g., casualty, destruction, emergency). An attempt was made to select words that were
unambiguous. For this reason, validation lexicons commonly for testing prejudice (Greenwald,
McGhee, & Schwartz, 1998; Greenwald, Nosek, & Banaji, 2003) and past word embedding
research (e.g., Caliskan, Bryson, & Narayanan, 2017), for example lists of insects and weapons,
were avoided. The words used in the validation lexicon are listed in Appendix C.
Once the validation lexicon was translated, word embeddings were obtained for
each word. Next, a search for exact duplicates was performed. In some languages, for example,
the word for man and men is the same. In other languages, semantically similar words like boy
and girl may be combined, resulting in a single word for both concepts, for example, child. Both
of these scenarios have a confounding effect on the results of the analysis and as such, duplicates
were removed.
Association was then calculated using cosine similarity. We hypothesized that a strong
validation of the method would result in a stronger association between known positive nouns
and positively valenced words on the one hand and known negative nouns and negatively
valenced words on the other. Moreover, these observations should be statistically significant.
Formally, a strong test of validation would satisfy both propositions:
cos(noun+, valence+) – cos(noun+, valence) > 0, p < 0.05
cos(noun, valence) – cos(noun, valence +) > 0, p < 0.05
58
Five languages, Cebuano, Japanese, Latin, Vietnamese, and Welsh failed validation
and were removed from the Wikipedia data set. Four languages, Cebuano, Hebrew, Latin, and
Vietnamese failed validation and were removed from the Common Crawl data set. The
remaining language data sets outlined in Table S1 were analyzed utilizing the method that
follows.
Table S1
Languages included in analysis after translation, classification, and validation.
Language
Genderedness
(Source)
Prejudice Analysis
Arabic
Gendered
(Haspelmath, 2005)
Wikipedia**, Common Crawl*
Armenian
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Azerbaijani
Genderless
(Simmons & Fennig, 2018)
Wikipedia, Common Crawl
Basque
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Belarusian
Gendered
(Simmons & Fennig, 2018)
Wikipedia***, Common
Crawl***
Bulgarian
Gendered
(Simmons & Fennig, 2018)
Wikipedia***, Common
Crawl***
Catalan
Gendered
(Simmons & Fennig, 2018)
Wikipedia, Common Crawl
Chinese (Mandarin)
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Czech
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia***, Common
Crawl***
Danish
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia*, Common Crawl
Dutch
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia**, Common Crawl**
English
Genderless
(Hellinger & Bußmann, 2001)
Wikipedia*, Common Crawl
Esperanto
Genderless
(Hellinger & Bußmann, 2001)
Wikipedia***, Common Crawl
Estonian
Genderless
Wikipedia*. Common Crawl
59
(Hellinger & Bußmann, 2001)
Finnish
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
French
Gendered
(Haspelmath, 2005)
Wikipedia***, Common
Crawl***
Galician
Gendered
(Simmons & Fennig, 2018)
Wikipedia, Common Crawl
Georgian
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
German
Gendered
(Haspelmath, 2005)
Wikipedia, Common Crawl
Greek
Gendered
(Haspelmath, 2005)
Wikipedia**, Common Crawl***
Hebrew
Gendered
(Corbett, 2013)
Wikipedia*
Hindi
Gendered
(Haspelmath, 2005)
Wikipedia**, Common Crawl***
Hungarian
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Indonesian
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Italian
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia, Common Crawl***
Japanese
Genderless
(Hellinger & Bußmann, 2001)
Common Crawl
Korean
Genderless
(Simmons & Fennig, 2018)
Wikipedia*, Common Crawl
Lithuanian
Gendered
(Simmons & Fennig, 2018)
Wikipedia, Common Crawl
Malay
Genderless
(Simmons & Fennig, 2018)
Wikipedia, Common Crawl
Norwegian
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia, Common Crawl
Persian
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Polish
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia**, Common Crawl***
Portuguese
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia, Common Crawl***
Romanian
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia***, Common
Crawl***
Russian
Gendered
(Haspelmath, 2005)
Wikipedia***, Common
Crawl***
60
Serbian
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia***, Common
Crawl***
Slovak
Gendered
(Simmons & Fennig, 2018)
Wikipedia***, Common
Crawl***
Slovenian
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia***, Common
Crawl***
Spanish
Gendered
(Haspelmath, 2005)
Wikipedia*, Common Crawl*
Swedish
Gendered
(Hellinger & Bußmann, 2001)
Wikipedia, Common Crawl
Tamil
Gendered
(Haspelmath, 2005)
Wikipedia, Common Crawl*
Thai
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Turkish
Genderless
(Haspelmath, 2005)
Wikipedia, Common Crawl
Ukrainian
Gendered
(Haspelmath, 2005)
Wikipedia, Common Crawl***
Urdu
Gendered
(Simmons & Fennig, 2018)
Wikipedia, Common Crawl
Welsh
Gendered
(Hellinger & Bußmann, 2001)
Common Crawl
*p < 0.05, **p < 0.01, ***p < 0.001.
Appendix B: Testing for Gender Prejudice
To identify gender prejudice in text corpora, we adapted a method used extensively
in social psychology to study the presence and influence of prejudice and stereotypes
(Greenwald et al., 1998, 2003). It suggests that if a target word is more strongly associated with a
positively than a negatively valenced word in a person’s memory, she will respond faster to the
pairing of the target word with a positive than a negative word. After employing the fastText
algorithm on the text corpora (Bojanowski, Grave, Joulin, & Mikolov, 2016; Grave, Bojanowski,
Gupta, Joulin, & Mikolov, 2018) a word is represented as a point in vector space with 300
dimensions. Given these dimensions, we can then find the level of similarity between two points
(words), as suggested by Caliskan et al. (2017). The resulting score captures the strength of
61
association between words. Therefore, we use semantic similarity between words as a measure of
association.
Formally, let X represent a set of female words and Y a set of male words. Let A be
the set containing positively valenced words and B the set containing negatively valenced words.
Cosine similarity is used to measure the similarity between two words where each word has an
associated vector. Thus, for target word x X and a valence word a A, Cos(x, a) = and x ·
a is the dot product of vectors x and a. Hence, we can calculate the similarity between x and each
of the members of A such that the meanaÎACos(x, a), denotes the average similarity between x
and all the positive words. Similarly, the estimated average similarity of x to all negative words
in set B would be given by meanbÎBCos(x, b).
The difference between meanaÎACos(x, a) and meanbÎBCos(x, b) provides a single
measure of relative similarity between gender word x and all positive and negative words.
Intuitively, such a difference will provide the net similarity to positive words. If meanaÎACos(x,
a) meanbÎBCos(x, b) > 0, then it demonstrates that x is more closely associated with positive
words than negative words. Conversely, if meanaÎACos(x, a) meanbÎBCos(x, b) < 0, it
demonstrates x is more closely associated with negative words. If S(x,A,B) = meanaÎACos(x, a)
meanbÎBCos(x, b) , then S(x,A,B) captures the sum of the net similarities for all members
of x. Similarly, S(y,A,B) captures the sum of the net similarities for all members of Y. The
main measure of prejudice would then be:
(Equation 1)
62
A positive value indicates the words in X are more closely associated with positive words than
the words in set Y. However, a negative value indicates words in Y are more closely associated
with positive words than those in X.
Permutation test. However, S(X,Y,A,B) is a single measure of relative similarity and
could thus be the result of random error. To estimate the probability that the obtained similarity
was not observed due to random allocations in X and Y, words are shuffled between groups. Put
another way, words from the original sets X and Y are randomly shuffled between groups to
create X* and Y* for the purposes of calculating S(X*,Y*,A,B). If the null hypothesis is true in such
a non-parametric permutation test, the result of S(X*,Y*,A,B) should be no different than
S(X,Y,A,B).
Formally, if (Xi,Yi) represents the potential random shuffling of words in X and Y,
then the probability of the observed score being the result of a random process will be:
(Equation 2)
Appendix C: Gender and Attribute Words
The Gender-specific lexicon was adapted from Bolukbasi, Change, Zou, Saligrama,
and Kalai (2016) and consists of 216 items separated by gender.
1. Male words: he, his, him, man, men, spokesman, himself, son, father, chairman, husband,
guy, boy, boys, brother, male, brothers, dad, sons, boyfriend, king, businessman,
grandfather, uncle, males, congressman, grandson, bull, businessmen, nephew, fathers,
lads, lion, gentleman, fraternity, bachelor, bulls, husbands, prince, colt, salesman, dude,
beard, councilman, gentleman, stepfather, monks, lad, sperm, testosterone, nephews,
daddy, fiancé, kings, dads, sir, stud, lions, czar, countrymen, penis, bloke, spokesmen,
63
suitors, monastery, brethren, prostate, schoolboy, brotherhood, stepson, uncles, monks,
viagra, paternity, suitor, macho, statesman, fathered, hubby, blokes, dudes, strongman,
grandsons, studs, godfather, boyhood, baritone, grandpa, boyfriends, countryman,
stallion, semen, fella, chap, salesmen, widower, beau, beards, handyman, horsemen,
fatherhood, princes, colts, fraternities, pa, fellas, councilmen, barbershop, fraternal,
prostate cancer, younger brother, ex boyfriend, twin brother, eldest son, estranged
husband, elder brother
2. Female words: her, she, women, woman, wife, mother, daughter, girls, girl,
spokeswoman, female, sister, herself, actress, mom, girlfriend, daughters, lady, sisters,
mothers, grandmother, deer, ladies, queen, wives, widow, bride, females, aunt, lesbian,
chairwoman, moms, maiden, granddaughter, niece, hers, princess, lesbians, actresses,
maiden, mare, fiancée, waitress, maternal, heroine, nieces, girlfriends, mistress, womb,
grandma, maternity, estrogen, widows, diva, nuns, nun, brides, housewife, menopause,
motherhood, stepmother, hostess, fillies, congresswoman, witch, sorority,
businesswoman, gal, schoolgirl, goddess, stepdaughter, uterus, mama, hens, hen,
mommy, grandmothers, feminism, heiress, queens, witches, aunts, granddaughters,
convent, vagina, maids, gals, housewives, obstetrics, councilwoman, matriarch, ma,
dowry, ballerina, ex girlfriend, estranged wife, ovarian cancer, teenage girl, twin sister
The positive and negative (pleasant and unpleasant) lexicon was adapted from Caliskan et al.
(2017).
64
Positive words: caress, freedom, health, love, peace, cheer, friend, heaven, loyal, pleasure,
diamond, gentle, honest, lucky, rainbow, diploma, gift, honor, miracle, sunrise, family,
happy, laughter, paradise, vacation
Negative words: abuse, crash, filth, murder, sickness, accident, death, grief, poison, stink,
assault, disaster, hatred, pollute, tragedy, divorce, jail, poverty, ugly, cancer, kill, rotten,
vomit, agony, prison
The warmth and competence lexicon was adapted from Rosenberg, Nelson, and Vivekananthan
(1968), Rudman, Greenwald, and McGhee (2001), Fiske, Cuddy, and Glick (2007), and Abele et
al. (2016).
Warmth words: friendly, good natured, sincere, warm, tolerant, caring, just, fair,
considerate, trustworthy, reliable, appreciative, honest, sentimental, happy, humorous,
sociable, supportive, nurture, nice, gentle, accepting, pleasant, loving, open
Competence words: capable, competent, confident, skillful, independent, competitive,
intelligent, clever, efficient, decisive, persistent, determined, industrious, imaginative,
potent, commanding, assertive, successful, bold, dynamic, logical, triumphant, dominant,
respectable, accomplished
The positive and negative validation lexicon was constructed by the authors.
Positive validation words: accomplishment, beauty, celebration, comfort, delight,
encouragement, fun, generous, genius, hero, holiday, humor, innovation, kiss, music,
nature, relief, safety, satisfaction, smile, success, sweetness, tranquility, victory, welcome
Negative validation words: arson, avalanche, blackout, blizzard, bushfire, casualty,
cyclone, destruction, drought, earthquake, emergency, erosion, fatality, flood, hailstorm,
65
hurricane, mishap, sandstorm, storm, tornado, tremor, tsunami, volcano, whirlpool,
windstorm
Appendix D: Range of prejudice across languages
Across the 45 languages, the means of each of the four comparisons in the Wikipedia data
set are reported in Table S2. The table shows average similarity of female (or male) words with
positively (or negatively) valenced words for each of the languages. The same relationships are
detailed in Table S3 for the Common Crawl data set.
Organizing the data in this way allows us to test the three-way interaction 2
(genderedness of languages: gendered vs. genderless) x 2 (gender words: male vs. female) x 2
(valence words: positive vs. negative). For the Wikipedia data set, this interaction is not
significant, F(7, 168) = 1.237, p > 0.05. However, for the Common Crawl data set, the three-way
interaction is significant, F(7, 172) = 6.086, p < 0.001. Decomposing this interaction across
gender words suggests that female words are more associated with positive words (M = 0.1816)
when a language is genderless than when a language is gendered (M = 0.1629; t = -2.9695, p <
0.01) even as negative words are relatively more associated with female words in genderless
languages (M = 0.1671) than gendered languages (M = 0.1557; t = 2.0203, p > 0.05). At the same
time, male words are associated with positive words in genderless (M = 0.1772) and gendered (M
= 0.1710; t = -0.8178, p > 0.05) languages to a relatively similar degree, but are much more
strongly associated with negative words in genderless (M = 0.1607) as opposed to gendered (M =
0.1480; t = -1.9681, p > 0.05) languages.
Table S2
Range of prejudice in the Wikipedia data set.
66
Language
Female-Negative
Female-Positive
Male-Negative
Male-Positive
Arabic
0.1848
0.1880
0.1772
0.1901
Armenian
0.2145
0.2177
0.2039
0.2069
Azerbaijani
0.1781
0.1929
0.1704
0.1815
Basque
0.2250
0.2291
0.2077
0.2113
Belarusian
0.2228
0.2115
0.1931
0.2005
Bulgarian
0.2065
0.1961
0.1921
0.1942
Catalan
0.2133
0.2237
0.2046
0.2156
Chinese
0.8852
0.8837
0.8781
0.8780
Czech
0.1969
0.1975
0.1790
0.1961
Danish
0.2255
0.2176
0.2020
0.2039
Dutch
0.2407
0.2353
0.2249
0.2296
English
0.2137
0.2220
0.1936
0.2100
Esperanto
0.2374
0.2185
0.2141
0.2113
Estonian
0.2078
0.2004
0.1857
0.1866
Finnish
0.2107
0.2391
0.1869
0.2096
French
0.2188
0.2102
0.1965
0.2105
Galician
0.1999
0.2089
0.1912
0.1954
Georgian
0.2226
0.2187
0.1969
0.1969
German
0.1882
0.2092
0.2304
0.2592
Greek
0.2102
0.2085
0.1956
0.2047
Hebrew
0.1803
0.1869
0.1603
0.1839
Hindi
0.2505
0.2457
0.2371
0.2448
Hungarian
0.2007
0.2168
0.1940
0.2134
Indonesian
0.2321
0.2277
0.2180
0.2216
Italian
0.2115
0.2230
0.2049
0.2227
Korean
0.2777
0.2864
0.2684
0.2657
Lithuanian
0.2126
0.2189
0.2007
0.2053
Malay
0.2325
0.2343
0.2219
0.2265
Norwegian
0.2151
0.2175
0.2032
0.2110
Persian
0.2389
0.2374
0.2281
0.2303
Polish
0.2071
0.2147
0.1842
0.2014
Portuguese
0.2068
0.2212
0.2078
0.2218
Romanian
0.2310
0.2197
0.2162
0.2218
Russian
0.2010
0.2076
0.1747
0.2022
Serbian
0.2547
0.2520
0.2317
0.2470
Slovak
0.2158
0.2103
0.1948
0.2060
Slovenian
0.2081
0.2071
0.1775
0.1896
Spanish
0.2091
0.2123
0.2022
0.2141
Swedish
0.2766
0.2783
0.2688
0.2724
Tamil
0.2000
0.2053
0.1877
0.1942
Thai
0.2323
0.2523
0.2223
0.2408
67
Turkish
0.1952
0.2041
0.1856
0.1966
Ukrainian
0.2131
0.2206
0.1878
0.2001
Urdu
0.2802
0.2764
0.2725
0.2713
Mean
(SD)
Mean
(SD)
Mean
(SD)
Mean
(SD)
Gendered
0.2171
(0.0243)
0.2187
(0.0219)
0.2035
(0.0261)
0.2146
(0.0242)
Genderless
0.2628
(0.1675)
0.2676
(0.1658)
0.2485
(0.1695)
0.2555
(0.1673)
Table S3
Range of prejudice in the Common Crawl data set.
Language
Female-Negative
Female-Positive
Male-Negative
Male-Positive
Arabic
0.1306
0.1519
0.1285
0.1578
Armenian
0.1689
0.1955
0.1646
0.1875
Azerbaijani
0.1451
0.1578
0.1298
0.1464
Basque
0.1343
0.1496
0.1235
0.1377
Belarusian
0.1631
0.1548
0.1420
0.1557
Bulgarian
0.1465
0.1450
0.1300
0.1605
Catalan
0.1762
0.1925
0.1840
0.1962
Chinese
0.1820
0.1867
0.1787
0.1887
Czech
0.1555
0.1487
0.1356
0.1582
Danish
0.1693
0.1772
0.1533
0.1641
Dutch
0.1671
0.1698
0.1680
0.1814
English
0.1454
0.1518
0.1383
0.1461
Esperanto
0.1863
0.1967
0.1723
0.1901
Estonian
0.1510
0.1607
0.1416
0.1461
Finnish
0.1791
0.2199
0.1758
0.2194
French
0.1676
0.1738
0.1570
0.2024
Galician
0.1571
0.1705
0.1583
0.1640
Georgian
0.1544
0.1757
0.1555
0.1769
German
0.1748
0.2070
0.1748
0.2255
Greek
0.1595
0.1655
0.1506
0.1801
Hindi
0.1379
0.1284
0.1284
0.1431
Hungarian
0.1553
0.1906
0.1451
0.1854
Indonesian
0.2064
0.2123
0.1980
0.2097
Italian
0.1540
0.1711
0.1537
0.1886
Japanese
0.1939
0.1918
0.1822
0.1805
Korean
0.1717
0.1661
0.1647
0.1570
Lithuanian
0.1625
0.1764
0.1683
0.1887
Malay
0.1746
0.1919
0.1712
0.1939
68
Norwegian
0.1742
0.1967
0.1718
0.1949
Persian
0.1657
0.1769
0.1521
0.1646
Polish
0.1499
0.1601
0.1470
0.1723
Portuguese
0.1481
0.1578
0.1506
0.1756
Romanian
0.1420
0.1469
0.1355
0.1628
Russian
0.1461
0.1477
0.1160
0.1600
Serbian
0.1613
0.1618
0.1488
0.1752
Slovak
0.1515
0.1487
0.1290
0.1469
Slovenian
0.1522
0.1538
0.1265
0.1451
Spanish
0.1450
0.1555
0.1482
0.1679
Swedish
0.1863
0.2029
0.1924
0.2153
Tamil
0.1504
0.1633
0.1448
0.1679
Thai
0.1760
0.1851
0.1868
0.2094
Turkish
0.1508
0.1774
0.1511
0.1723
Ukrainian
0.1829
0.1745
0.1676
0.1866
Urdu
0.1383
0.1475
0.1325
0.1479
Welsh
0.1088
0.1106
0.1010
0.1028
Mean
(SD)
Mean
(SD)
Mean
(SD)
Mean
(SD)
Gendered
0.1558
(0.01665)
0.1629
(0.0212)
0.148
(0.0208)
0.1709
(0.0246)
Genderless
0.1671
(0.0194)
0.1816
(0.0199)
0.1607
(0.0209)
0.1772
(0.0246)
Appendix E: Additional Analysis for Gender Prejudice
It is worth noting that languages can belong to different language families (e.g.,
Indo-European, Sino-Tibetan). Therefore, one could argue that the observed difference between
gendered versus genderless languages is not because of genderedness but because of the
language family. To address this account, we conducted a Hierarchical Linear model (HLM) with
language family as the second-level variable. We classified the 45 languages using data from
Ethnologue (Simmons & Fennig, 2018; see Table S4). An analysis using a random-intercept
HLM model that controls for variance in each language family as well as across, showed that
genderedness still has a significant influence on net gender prejudice in both the Wikipedia and
Common Crawl data. For the Wikipedia data language family had a positive influence on net
69
prejudice (
𝛽
=0.0096, p < 0.001) and classification as a genderless language reduced the net
prejudice (
𝛽
=-0.0074, p < 0.01). The analysis in the Common Crawl data also indicated that
language family had a positive influence on net prejudice (
𝛽
=0.0158, p < 0.001) and
classification as a genderless language reduced the net prejudice (
𝛽
=-0.0137, p < 0.001).
Table S4
Languages classified by their macro-family.
Language
Language Family
Arabic
Afro-Asiatic
Armenian
Indo-European
Azerbaijani
Turkic
Basque
Isolate
Belarusian
Indo-European
Bulgarian
Indo-European
Catalan
Indo-European
Chinese (Mandarin)
Sino-Tibetan
Czech
Indo-European
Danish
Indo-European
Dutch
Indo-European
English
Indo-European
Esperanto
International Auxiliary Languages
Estonian
Uralic
Finnish
Uralic
French
Indo-European
Galician
Indo-European
Georgian
Caucasian
German
Indo-European
Greek
Indo-European
Hebrew
Afro-Asiatic
Hindi
Indo-European
Hungarian
Uralic
Indonesian
Austronesian
Italian
Indo-European
Japanese
Japonic
Korean
Koreanic
70
Lithuanian
Indo-European
Malay
Austronesian
Norwegian
Indo-European
Persian
Indo-European
Polish
Indo-European
Portuguese
Indo-European
Romanian
Indo-European
Russian
Indo-European
Serbian
Indo-European
Slovak
Indo-European
Slovenian
Indo-European
Spanish
Indo-European
Swedish
Indo-European
Tamil
Dravidian
Thai
Tai-Kaidai
Turkish
Turkic
Ukrainian
Indo-European
Urdu
Indo-European
Welsh
Indo-European
One could also argue that there may be a systematic variation such that prosperous
countries are more likely to have genderless language while less prosperous countries have
gendered languages or vice versa, and that prosperity of the country in which a language is
spoken affects the pattern of our findings. However, the content contributed online (e.g., on
Wikipedia) in a certain language can come from people in different countries (e.g., people in any
country can contribute to online English, Spanish, or French content if they speak that language).
As a result, we do not conclusively observe such a systematic pattern in our data.
Finally, one could argue that there may be specific features of the region where
languages are spoken that could affect gender prejudice over and above the genderedness of
languages. In order to address such a concern we ran our analysis controlling for geographic,
cultural, and demographic covariates that have been shown to influence gender equality (Prewitt-
71
Freilino, Caswell, & Laakso, 2012; United Nations Development Programme, 2017; World
Economic Forum, 2017).
We used the Global Gender Gap index developed by World Economic Forum
(2017), the Corruption Perceptions Index developed by Transparency International (2016), the
Religious Diversity Index developed by Pew Research Center (2014), the STC Climate Index
compiled from public weather data (Tajick, 2018), and the Human Development Index
developed by the United Nations Development Programme (2017). Since most global indices are
calculated per country, not per language, we utilized data from Wikipedia outlining the relative
contribution of residents from each country to each language project (Wikimedia Foundation,
2018). Country-level index values were then weighted and averaged for each language, based on
that country’s relative contribution to the Wikipedia language project. This same distribution was
used as a proxy to analyze both Wikipedia and Common Crawl data. Covariate values were
centered to make interpretation of coefficients easier.
For the Wikipedia data, a linear regression indicated that the predictor of
genderedness significantly predicted gender prejudice for each language. The results indicated
that the predictor explained 21.3% of the total variance (R2 = 0.2131, F(1,42) = 11.37, p < 0.01).
Gendered languages had more gender prejudice than genderless languages (β = 0.0074, p <
0.01). A more complex model including the covariates of gender parity, corruption, religious
diversity, development, and climate did not improve predictive ability (R2 = 0.2132, F(6,34) =
1.535, p > 0.05). Therefore, even after controlling for covariates gendered languages displayed
more gender prejudice than genderless languages.
72
A similar pattern of results emerged for the Common Crawl data. A linear regression
indicated a significant effect of genderedness on gender prejudice; it explained 30.2% of the total
variance (R2 = 0.3017, F(1,43) = 18.57, p < 0.001). Gendered language had more gender
prejudice than genderless languages (β = 0.0137, p < 0.001). In the Common Crawl data, one
covariate, corruption, along with genderedness improved the predictive ability of the model (R2 =
0.5128, F(6,35) = 6.14, p < 0.001). Specifically, genderedness (β = 0.0142, p < 0.001) and
corruption (β = -0.0006, p < 0.01)1 predicted gender prejudice.
In sum, the regression model results (see Table S5) suggest that the presence of
genderedness alone is responsible for between 20% and 30% of the variance in observed gender
prejudice across 45 languages in two different data sets, even after accounting for geographic,
demographic, and cultural confounds.
Table S5
Summary of the regression analysis including geographic, demographic and cultural control
variables.
Wikipedia
Gender Model
Wikipedia
Complete Model
Common Crawl
Gender Model
Common Crawl
Complete Model
Variable
β
(SE)
β
(SE)
β
(SE)
β
(SE)
Intercept
0.0022
(0.0018)
-0.0068
(0.0248)
0.0021
(0.0025)
-0.0130
(0.0313)
Genderedness
0.0074**
(0.0022)
0.0075*
(0.0028)
0.0137***
(0.0032)
0.0142***
(0.0034)
Global Gender Gap
-0.0090
(0.0420)
0.0473
(0.0494)
Corruption Perceptions
-0.0001
(0.0309)
-0.0006**
(0.0002)
Religious Diversity
0.0005
(0.0010)
0.0019
(0.0012)
73
STC Climate Index
0.0008
(0.0121)
-0.0121
(0.0153)
Human Development Index
0.0104
(0.0310)
0.0250
(0.0388)
R2
0.2131
0.2132
0.3017
0.5128
F
11.37**
1.5350
18.57***
6.14***
* p < 0.05, ** p < 0.01, *** p < 0.001.
Appendix F: Additional Analysis for Gender Stereotype
A similar analysis was conducted for the stereotypic associations of gender with
warmth and competence. Since warmth and competence are not binary variables like positive
and negative, the two constructs are considered separately. A random-intercept HLM that
controls for the variance in each language family as well as across each language showed that the
observed patterns held. For the Wikipedia data, language family did not have a significant
influence on associations between gender and warmth words (p = 0.1693), but the presence of
genderedness (β = 0.0076, p < 0.05) had a positive influence on the difference in association
between gender words and warmth words. The analysis in the Common Crawl data again
indicated that language family did not have an influence on associations between gender and
warmth words (p = 0.451); however, genderedness did have a significant positive influence on
the difference in association between gender words and warmth words (β = 0.0134, p < 0.05).
For the Wikipedia data, both language family (β = 0.0094, p < 0.001) and genderedness (β =
0.0079, p < 0.05) had a significant influence on the difference between the association of gender
words and competence. Similarly, for the Common Crawl data, both language family (β =
0.0079, p < 0.01) and genderedness (β = 0.0168, p < 0.001) had a significant influence on the
difference between the association of gender words and competence. This shows that even after
taking onto account language family, the influence of genderedness of language persists.
74
Again, to account for the influence of covariates, we used the Global Gender Gap
index developed by World Economic Forum (2017), the Corruption Perceptions Index developed
by Transparency International (2016), the Religious Diversity Index developed by Pew Research
Center (2014), the STC Climate Index compiled from public weather data (Tajick, 2018), and the
Human Development Index developed by the United Nations Development Programme (2017),
this time with the results for the warmth-competence constructs weighted by the proportional
contribution of each country to each language to Wikipedia, to account for possible cultural and
geographic covariates (Prewitt-Freilino et al., 2012).
For the Wikipedia data, a linear regression indicated that the predictor of
genderedness significantly predicted gender stereotypes associated with warmth. The results
indicated that the predictor explained 16.8% of the total variance (R2 = 0.168, F(1, 42) = 8.748, p
< 0.01). Gendered languages showed a greater difference in association between gender words
and warmth words (β = 0.0071, p < 0.01). A more complex model including the covariates of
gender parity, corruption, religious diversity, development, and climate improved predictive
ability (R2 = 0.4681, F(6, 34) = 4.986, p < 0.001), with a similar relationship between gendered
languages and the difference in association between gender and warmth words (β = 0.0072, p <
0.01) and none of the other covariates showing a significant influence (see Table S6).
The same analysis was conducted for the Common Crawl data. A linear regression
indicated a significant effect of genderedness on gender stereotypes; it explained 34% of the total
variance (R2 = 0.3404, F(1, 43) = 22.19, p < 0.001). Gendered languages showed a greater
difference in the association between gender words and warmth words (β = 0.0134, p < 0.001).
With the more complex model suggested by Prewitt-Freilino et al. (2012), improved predictive
75
ability (R2 = 0.5658, F(6, 35) = 7.6, p < 0.001), with a similar relationship between gendered
languages and the difference in association between gender and warmth words (β = 0.0143, p <
0.001) and none of the other covariates showed a significant influence (see Table S6).
Table S6
Summary of the regression analysis for the stereotype of warmth, including geographic,
demographic and cultural control variables.
Wikipedia
Warmth
Gender Model
Wikipedia
Warmth
Complete Model
Common Crawl
Warmth
Gender Model
Common Crawl
Warmth
Complete Model
Variable
β
(SE)
β
(SE)
β
(SE)
β
(SE)
Intercept
-0.0032
(0.002)
-0.0212
(0.0205)
-0.0017
(0.0023)
0.0201
(0.0252)
Genderedness
0.0071**
(0.0024)
0.0072**
(0.0025)
0.0134***
(0.0029)
0.0143***
(0.003)
Global Gender Gap
-0.0415
(0.0359)
-0.0166
(0.0417)
Corruption Perceptions
-0.0001
(0.0001)
-0.0003
(0.0002)
Religious Diversity
-0.0001
(0.0001)
0.0001
(0.001)
STC Climate Index
-0.0001
(0.0107)
-0.0179
(0.0131)
Human Development Index
-0.0275
(0.0252)
-0.0152
(0.0309)
R2
0.168
0.4681
0.3404
0.5128
F
8.478**
4.986***
22.19***
6.14***
* p < 0.05, ** p < 0.01, *** p < 0.001.
The analysis was repeated for competence words in each data set. For the Wikipedia
data set, a linear regression indicated genderedness was a significant predictor of gender
stereotypes associated with competence. The results indicated that the predictor explained 14%
76
of the total variance (R2 = 0.1435, F(1, 42) = 7.034, p < 0.05). Gendered languages showed a
greater difference in association between gender words and competence words (β = 0.0079, p <
0.05). A more complex model including the covariates of gender parity, corruption, religious
diversity, development, and climate improved predictive ability (R2 = 0.3076, F(6, 34) = 2.517, p
< 0.05), with a similar relationship between gendered languages and the difference in association
between gender and competence words (β = 0.0085, p < 0.05) and none of the covariates
showing significant influence (see Table S6). For the Common Crawl data, a linear regression
indicated a significant effect of genderedness on gender stereotypes of competence; it explained
14% of the total variance (R2 = 0.1435, F(1, 42) = 7.034, p < 0.05). Gendered languages showed
a greater difference in the association between gender words and competence words (β = 0.0079,
p < 0.05). The more complex model improved predictive ability (R2 = 0.3076, F(6, 34) = 2.517, p
< 0.05), with a similar relationship between gendered languages and the difference in association
between gender and competence words (β = 0.0085, p < 0.05) and none of the other covariates
showing a significant influence (see Table S7).
Table S7
Summary of the regression analysis for the stereotype of warmth, including geographic,
demographic and cultural control variables.
Wikipedia
Competence
Gender Model
Wikipedia
Competence
Complete Model
Common Crawl
Competence
Gender Model
Common Crawl
Competence
Complete Model
Variable
β
(SE)
β
(SE)
β
(SE)
β
(SE)
Intercept
0.0094
(0.0024)
-0.0292
(0.0285)
0.0094
(0.0024)
-0.0292
(0.0285)
77
Genderedness
0.0079 **
(0.0029)
0.0085*
(0.0035)
0.0079*
(0.0029)
0.0085*
(0.0035)
Global Gender Gap
-0.0233
(0.0499)
-0.0232
(0.0499)
Corruption Perceptions
-0.0003
(0.0002)
-0.0003
(0.0002)
Religious Diversity
-0.0019
(0.0012)
-0.0019
(0.0012)
STC Climate Index
-0.0226
(0.0149)
-0.0226
(0.0149)
Human Development Index
0.0595
(0.0349)
0.0595
(0.03499)
R2
0.1435
0.3076
0.1435
0.3076
F
7.034*
2.517*
7.034*
2.517*
* p < 0.05, ** p < 0.01, *** p < 0.001.
78
References
Abele, A. E., Hauke, N., Peters, K., Louvet, E., Szymkow, A., & Duan, Y. (2016). Facets of the
fundamental content dimensions: Agency with competence and assertiveness—communion with
warmth and morality. Frontiers in Psychology, 7.
Bojanowski, P., Grave, E., Joulin, A., & Mikolov, T. (2016). Enriching word vectors with subword
information. ArXiv:1607.04606 [Cs]. Retrieved from http://arxiv.org/abs/1607.04606
Bolukbasi, T., Chang, K.-W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer
programmer as woman is to homemaker? Debiasing word embeddings. In D. D. Lee, M.
Sugiyama, U. V. Luxburg, I. Guyon, & R. Garnett (Eds.), Advances in neural information
processing systems (Vol. 29, pp. 4349–4357). Retrieved from http://papers.nips.cc/paper/6228-
man-is-to-computer-programmer-as-woman-is-to-homemaker-debiasing-word-embeddings.pdf
Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language
corpora contain human-like biases. Science, 356(6334), 183–186.
Corbett, G. G. (1991). Gender. New York, NY: Cambridge University Press.
Fiske, S. T., Cuddy, A. J. C., & Glick, P. (2007). Universal dimensions of social cognition: Warmth
and competence. Trends in Cognitive Sciences, 11(2), 77–83.
Grave, E., Bojanowski, P., Gupta, P., Joulin, A., & Mikolov, T. (2018). Learning word vectors for
157 languages. ArXiv:1802.06893 [Cs]. Retrieved from http://arxiv.org/abs/1802.06893
Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. K. (1998). Measuring individual differences in
implicit cognition: The Implicit Association Test. Journal of Personality and Social Psychology,
74(6), 1464–1480.
79
Greenwald, A. G., Nosek, B. A., & Banaji, M. R. (2003). Understanding and using the Implicit
Association Test: I An improved scoring algorithm. Journal of Personality and Social
Psychology, 85(2), 197–216.
Haspelmath, M. (2005). World Atlas of Language Structures. Oxford, UK: Oxford University Press.
Hellinger, M., & Bußmann, H. (Eds.). (2001). Gender across languages: The linguistic representation
of women and men. Philadelphia, PA: John Benjamins Publishing Company.
Pew Research Center. (2014). Religious diversity index scores by country. Retrieved from
http://www.pewforum.org/2014/04/04/religious-diversity-index-scores-by-country/
Prewitt-Freilino, J. L., Caswell, T. A., & Laakso, E. K. (2012). The gendering of language: A
comparison of gender equality in countries with gendered, natural gender, and genderless
languages. Sex Roles, 66(3–4), 268–281.
Rosenberg, S., Nelson, C., & Vivekananthan, P. S. (1968). A multidimensional approach to the
structure of personality impressions. Journal of Personality and Social Psychology, 9(4), 283–
294.
Rudman, L. A., Greenwald, A. G., & McGhee, D. E. (2001). Implicit self-concept and evaluative
implicit gender stereotypes: Self and ingroup share desirable traits. Personality and Social
Psychology Bulletin, 27(9), 1164–1178.
Simmons, G. F., & Fennig, C. D. (2018). Ethnologue: Languages of the World. Retrieved from
https://www.ethnologue.com/
Tajick, S. (2018). STC climate index. Retrieved from http://globalresidenceindex.com/hnwi-
index/climate-index/
80
Transparency International. (2016). Corruption perceptions index. Retrieved from
https://www.transparency.org/news/feature/corruption_perceptions_index_2016
United Nations Development Programme. (2017). Human development report 2016: Human
development for everyone. New York, NY: United Nations Development Programme.
Wikimedia Foundation. (2018). Wikimedia traffic analysis report: Page views per Wikipedia
language. Retrieved from https://stats.wikimedia.org/archive/squid_reports/2018-
05/SquidReportPageViewsPerLanguageBreakdown.htm
World Economic Forum. (2017). The global gender gap report: 2017. Geneva: World Economic
Forum.
81
Footnotes
1 Note that the corruption index is scaled such that a number closer to zero indicates more
public corruption and a number closer to one indicates less public corruption (Transparency
International, 2016).
... In the vast majority of social cognition and perception theories and literature two dimensions appear, namely warmth-communion and competence-agency dimensions (Sainz et al. 2021), based in agency-communion literature and SCM literature. It has been demonstrated that these dimensions occur across regions and cultures DeFranza et al. 2020;Durante et al. 2017). ...
... In comparable fashion, the SCM builds on the agency communion literature stating that all interpersonal impressions and group stereotypes form along two dimensions, competence and warmth (Fiske et al. 2002), warmth is comparable to communion and consist of traits such as sincere, helpful, good natured, tolerant and warm, while competence is comparable to agency and consists of traits such as independent, competent, intelligent, confident and competitive (Awale et al. 2019;Brambilla et al. 2011;Caprariello et al. 2009;Cuddy et al. 2009;DeFranza et al. 2020;Fiske et al. 2002;Fiske 2018;Kervyn et al. 2013;Nejat et al. 2020). The two theories are often used interchangeably. ...
... Both in agency and communion literature, SCM literature and general stereotype literature one sees a tendency for warmth (communal traits) to be attributed to women, while competence (agency traits) are attributed to men (Awale et al. 2019;Brambilla et al. 2011;Caprariello et al. 2009;Conway et al. 1996;Cuddy et al. 2009;DeFranza et al. 2020;Eagly et al. 2000;Eagly and Karau 2002;Ellemers 2018;Fiske 2018;Fiske et al. 2002;Kevryn et al. 2013;Kurt et al. 2011;McDonnell and Baxter 2019;Nejat et al. 2020;Smith et al. 2018). The stereotypes that arise based on gender are prone to be high in one dimension; warmth (communion) or competence (agency) Fiske et al. 1999). ...
Article
Full-text available
Chatbots have in recent years increasingly been used by organizations to interact with their customers. Interestingly, most of these chatbots are gendered as female, displaying stereotypical notions in their avatars, profile pictures and language. Considering the harmful effects associated with gender-based stereotyping at a societal level—and in particular the detrimental effects to women—it is crucial to understand the effects of such stereotyping when transferred and perpetuated by chatbots. The current study draws on the Stereotype Content Model (SCM) and explores how the warmth (high vs. low) of a chatbot’s language and the chatbot’s assigned gender elicit stereotypes that affect the perceived trust, helpfulness, and competence of the chatbot. In doing so, this study shows how established stereotype theory can be used as a framework for human-machine communication research. Moreover, its results can serve as a foundation to explore ways of mitigating the perpetuation of stereotyping and bring forward a broader discussion on ethical considerations for human-machine communication.
... Due to the complexity of the space, no individual dimension of a word vector is interpretable. However, studying differences between vectors can yield remarkable insights regarding the semantic meaning of words and contribute to the understanding of gender stereotypes (31)(32)(33). In fact, the study of analogical reasoning using word vectors has highlighted the inherent gender biases in language, showing, for instance, that historically "man is to computer programmer as woman is to homemaker" (34, p. 1). ...
... Methodologically, we contribute to the nascent literature of using word embeddings to quantify gender stereotypes and social change (32,33,46). Recent designs have used word embedding models trained on large corpora of historical text to chart how the understanding of concepts has changed over time. ...
... Estimation of Semantic Associations between Vectors. In line with the canon of NLP, we assume that the dimensions of word vectors embody semantic meaning (27,(31)(32)(33)(34)67). This implies that the similarity between two vectors reflects how closely associated the two words represented by these vectors are, in terms of meaning (67). ...
Article
Full-text available
Significance Gender inequality has been deemed the “greatest human rights challenge of our time” by the United Nations, and scholars across numerous disciplines agree that gender stereotypes represent a primary way by which this inequality is maintained. Yet changing stereotypes in a systemic, enduring way is extremely difficult. This is at least in part because stereotypes are transmitted and perpetuated through the language societies and organizations use to describe women, especially those in leadership roles. Here, we show that hiring women into leadership positions is associated with organizations characterizing women in more leadership-congruent, agentic ways. This shift mitigates a critical barrier to women’s progression in organizations and society: the incongruence of what it means to be a woman and a leader.
... To this end, there are several reported cases of advanced language models seeming to exhibit potentially harmful biases. For example, models have shown that word embeddings for African American names are more likely to be associated with unpleasant terms than European American names (Caliskan et al., 2017) and that women are globally linked more closely to negatively valenced words than men (DeFranza et al., 2020;Garg et al., 2018). It has also been shown that models often favor people according to their age (e.g., biased toward the language of older adults; Hovy and Sogaard, 2015) and reflect the economic status and education of speakers (e.g., more likely to be trained on higher education and income samples; Graesser et al., 2014;Mota et al., 2018;Sultana et al., 2020). ...
... It has also been shown that models often favor people according to their age (e.g., biased toward the language of older adults; Hovy and Sogaard, 2015) and reflect the economic status and education of speakers (e.g., more likely to be trained on higher education and income samples; Graesser et al., 2014;Mota et al., 2018;Sultana et al., 2020). Failure to account for these issues can lead to undesirable outcomes in health care contexts (Cirillo et al., 2020;Obermeyer et al., 2019), hiring and evaluations in organizations and academia (Bertrand and Mullainathan, 2004;Harber et al., 2019), and efforts to reduce societal biases (DeFranza et al., 2020;Sap et al., 2014;Sun et al., 2019). In clinical contexts, being unaware of biases can result in false assumptions about clients that have negative treatment consequences. ...
Article
Technological advances in artificial intelligence and natural language processing have increased efficiency of assessing speech content and speech organization in schizophrenia. Despite these developments, there has been little focus on the psychometrics of these approaches. Using two common assessments, the current study addressed this gap by: 1) measuring test-retest reliability; and 2) assessing whether speech content and/or speech organization generalize across demographics. To test these aims, we examined psychometric properties of the Linguistic Inquiry Word Count (LIWC), a speech content measure, and the Coh-Metrix, a speech organization measure. Across baseline to six month (n = 101) and baseline to one year (n = 47) narrative speech samples, we generally observed fair reliability for speech content measures and fair to good reliability for speech organization measures. Regarding demographics, multiple speech indices varied by race, income, and education. The lack of excellent reliability scores for speech indices holds important implications for examining speech variables in clinical trials and highlights the dynamic nature of speech. This work illustrates the importance of designing speech content and speech organization measures with external validity across demographic factors. Future studies examining speech in schizophrenia should account for potential biases against demographic groups introduced by linguistic analysis tools.
... findings are consistent with the recent empirical evidence supporting the weak version of the Sapir-Whorf hypothesis. We show that the grammatical gender system of a language can influence speakers' judgments of the human-nonhuman divide, and in turn their anthropomorphism tendencies, which confirms and extends research on the effects of grammatical gender on object descriptions (Boroditsky et al., 2003), categorization (Cubelli et al., 2011;Sera et al., 1994), and attitudes, intentions, and behaviors (DeFranza et al., 2020;Mecit et al., 2021;Prewitt-Freilino et al., 2012). Consistent with the previous studies investigating the relativity thesis with bilingual participants (Kousta et al., 2008), we found that French-English bilingual speakers' answers differed in each of their languages (Study 1). ...
... For example, given that grammatical gender effects are more likely to occur in languages with only two grammatical genders (Maciuszek et al., 2019), such as French and Spanish (Sera et al., 1994), one avenue for future research is to test the generalizability of the effect in languages with more than two grammatical genders. That said, recent research investigating the effects of grammatical gender systems on gender prejudice (DeFranza et al., 2020) and gender equality (Prewitt-Freilino et al., 2012) compared across multiple countries (45 and 101, respectively). Thus, the generalizability of the effect, at least in terms of gender equality and prejudice, seems to be firmly established. ...
Article
When English speakers anthropomorphize animals or objects, they refer to such entities using human pronouns (e.g., he or she instead of it). Unlike English, which marks gender only for humans, gendered languages such as French grammatically mark gender not only for humans but also for nonhumans. Research has shown that in gendered languages, although gender marking of nonhuman nouns is semantically arbitrary, people ascribe male and female properties to nonhuman entities consistent with their grammatical gender. Because grammatical gender conveys human-related properties, we question whether grammatically gender-marking nonhumans may elicit anthropomorphism tendencies. Across six studies, we show that gender marking of nonhuman nouns in gendered languages influences the way individuals mentally represent these entities and increases their anthropomorphism tendencies. We demonstrate the effects both by comparing anthropomorphism as a function of natural differences in languages with French-English bilinguals (Study 1) and by training native English speakers to use gender marking for nonhuman nouns as speakers of gendered languages do (Study 2). The following studies further demonstrate the effects within the French language by measuring (Study 3a) and manipulating (Studies 3b and 4) the salience of gender markings of nonhuman nouns. In Study 5 (preregistered), we replicate our basic finding and establish grammatical gender as an important linguistic element in shaping French speakers' anthropomorphism tendencies. We discuss the findings and the limitations in the culture-language-cognition triad and layout their implications for the debate on the extent to which language can mediate categorical and perceptual judgments. (PsycInfo Database Record (c) 2022 APA, all rights reserved).
... Anthropocentric linguistics pays great importance to the study of the human factor in languages (DeFranza, Mishra, & Mishra, 2020;Fishman, 2019;Kuzembayeva et al., 2019;Sealey, 2018). Globalization and the intensification of intercultural communication contribute to the unification of language processes and an increase in international vocabulary while preserving ethno-cultural markings (Al Rawashdekh, 2019;Anderson, 2020;Egoreichenko, 2018). ...
Article
Full-text available
A cross-cultural investigation of lacunae contributes to successful intercultural communication, helping in achieving mutual understanding between cultures and solving communication problems between different nations. This study is aimed at identifying and classifying lexical lacunae in the kinship terminologies of the Russian and Kazakh languages. The comparative study reveals semantic, cultural, psychological, evaluative, and aesthetic factors. The material of the study comprises the terms of kinship, selected by the method of continuous sampling from lexicographic sources of the Kazakh and Russian languages, with a total volume of 300 units. Interlanguage lacunae were described from the perspective of comparative structural-semantic analysis. The analysis of kinship terms in the Russian and Kazakh languages showed that there are more similarities than previously thought. Significant differences in Russian and Kazakh kinship terminology were manifested in connection with the allocation of the seme "gender correlation", i.e., the name depends on the gender of the person concerning whom it was used, as well as the seme "age concerning the speaking person". Analysis of Kazakh kinship terms that act as lacunae in the Russian language and their classification in terms of semantics, structure, etymology and cultural components was not considered an object of research until now.
... Language has the power to shape and modify the way people think and perceive the world (DeFranza et al., 2020). Language indeed is far more influential and powerful when it is part of a social act. ...
Chapter
The human experience in times of crisis is a determinative indicator for the future wellbeing of generations. The lack of empathy and inactive emotional intelligence through all forms of linguistic conduct cause miscommunication and misconduct, which severely underestimates the intellectual potential of human beings. In a world of diversity, emotional intelligence and empathic linguistic power are crucial indicators of civilization and enlightenment. Given a richer understanding of the relationship between empathy and emotional intelligence from a sociolinguistic perspective, this study discusses the significance of including emotional intelligence and empathy in educational and intellectual programs. This study is the framework through which the empathic linguistic power within a society could be a determining power for crisis management and wellbeing at times of turmoil.
Article
Resumo: Os debates sobre o uso de formas não binárias incluem a possibilidade ou não de alterar regras existentes ou adicionar novas formas ao conjunto de normas da língua que viabilizam a expressão de identidades e a possibilidade de existências. No entanto, o uso de formas não binárias é visto por muitas pessoas como algo que transgride a norma culta ou modifica a língua como um todo, apesar de as línguas vivas passarem por constantes mudanças. A sala de aula de língua pode ou fomentar ou impossibilitar a expressão de identidades que escapam a sistemas rígidos, dependendo da visão de quem leciona. Este artigo discorre a respeito da visão de profissionais de língua portuguesa sobre o possível uso de linguagem inclusiva em sala de aula. Os dados obtidos através da aplicação de um questionário anônimo revelam a familiaridade de profissionais de língua portuguesa (n=55) com questões relacionadas à não binariedade, bem como dificuldades e preocupações em relação à implementação de linguagem inclusiva em aula. Os resultados indicam que a maioria das pessoas que participaram da pesquisa já tinham ouvido falar sobre linguagem neutra/não binária (93%) e conheciam pessoa não binária (74,5%). A maior parte do grupo (72%) concordava com o acréscimo de formas não binárias ao português, mas um número menor se disse confortável com o seu uso em geral (64%) e em materiais didáticos (57%). Dentre as preocupações expressas por profissionais, destaca-se a necessidade de desenvolvimento profissional. Disponível em https://www.portugueselanguagejournal.com/
Chapter
Hate speech continues to be an issue of key social significance, yet while its lexical and discursive aspects have been widely studied, its grammatical traits have been hitherto overlooked. This book seeks to address this gap by bringing together a global team of scholars to explore the morphosyntactic features of hateful and aggressive discourse. Drawing on thirteen diverse cross-linguistic case studies, it reveals how hate is expressed in political discourse, slang, and social media, and towards a range of target groups relating to gender, sexual orientation, and ethnic identity. Based on ideas from functional and cognitive linguistics, each thematic part demonstrates how features such as morphology, word formation, pronoun use, and syntactic structures are manipulated for the purpose of expressing hostility and hate. An innovative approach to an age-old problem, this book is essential reading for researchers and students of hate speech and verbal aggression.
Article
Over 5 days at the Nag’s Head Conference Center, USA in 1987, social and cross-cultural psychologists discussed what would be required if research relating to culture were to gain greater attention from psychology in general, and in particular from what was perceived at the time as its mainstream. The criteria for gaining greater credibility laid down by three leading social psychologists proved daunting in relation to the cross-cultural work presented at the meeting but subsequently inspired cross-culturalists to “raise their game.” In this paper, we describe these crucial challenges and how they have been addressed more recently by cross-cultural psychologists. We assess the extent to which studies focused on cultural differences are now thoughtfully represented in social, personality, and organizational psychology by briefly surveying the content of a single year’s issues of the Journal of Personality and Social Psychology, Journal of International Business Studies, and the Journal of Personality in relation to the concurrent content of the Journal of Cross-Cultural Psychology. We identify the perils of assimilation to psychology in general by diluting the concept of culture and by tyrannizing research with over-specified criteria of statistical rectitude. We also identify studies published in top-rated journals that have nonetheless advanced our field. We reiterate the need for defensible measures of cultural difference and methods for identifying and examining them as a basis for multi-level explanations of cultural effects and cultural change. We conclude by proposing a gold standard for assaying cross-cultural studies of psychological processes and outcomes.
Article
Full-text available
Most theories of emotion describe a crucial role for interoceptive accuracy, the perception of the body’s internal physiological signals, in emotional experience. Despite support for interoceptive accuracy’s role in emotion, findings of gender differences in emotional and interoceptive processing are incompatible with theory; women typically show poorer interoceptive accuracy, but women often outperform men on measures of emotional processing and recognition. This suggests a need to re-evaluate the relationship between interoceptive accuracy and emotion considering sex and gender. Here we extend Pennebaker and Roberts’ (1992) theory of gender differences in the use of interoceptive signals for emotional experience, proposing that language socialisation may result in gender differences in the propensity to label internal state changes as physiological or emotional, respectively. Despite outstanding questions concerning the fractionation of interoceptive and emotional domains, this theory provides a plausible explanation for seemingly incompatible findings of gender differences in interoceptive and emotional abilities.
Article
Full-text available
For nearly a century social scientists have sought to understand left-right ideological differences in values, motives, and thinking styles. Much progress has been made, but-as in other areas of research-this work has been criticized for relying on small and statistically unrepresentative samples and the use of reactive, self-report measures that lack ecological validity. In an effort to overcome these limitations, we employed automated text analytic methods to investigate the spontaneous, naturally occurring use of language in nearly 25,000 Twitter users. We derived 27 hypotheses from the literature on political psychology and tested them using 32 individual dictionaries. In 23 cases, we observed significant differences in the linguistic styles of liberals and conservatives. For instance, liberals used more language that conveyed benevolence, whereas conservatives used more language pertaining to threat, power, tradition, resistance to change, certainty, security, anger, anxiety, and negative emotion in general. In 17 cases, there were also significant effects of ideological extremity. For instance, moderates used more benevolent language, whereas extremists used more language pertaining to inhibition, tentativeness, affiliation, resistance to change, certainty, security, anger, anxiety, negative affect, swear words, and death-related language. These research methods, which are easily adaptable, open up new and unprecedented opportunities for conducting unobtrusive research in psycholinguistics and political psychology with large and diverse samples. (PsycINFO Database Record (c) 2020 APA, all rights reserved).
Article
Full-text available
Distributed word representations, or word vectors, have recently been applied to many tasks in natural language processing, leading to state-of-the-art performance. A key ingredient to the successful application of these representations is to train them on very large corpora, and use these pre-trained models in downstream tasks. In this paper, we describe how we trained such high quality word representations for 157 languages. We used two sources of data to train these models: the free online encyclopedia Wikipedia and data from the common crawl project. We also introduce three new word analogy datasets to evaluate these word vectors, for French, Hindi and Polish. Finally, we evaluate our pre-trained word vectors on 10 languages for which evaluation datasets exists, showing very strong performance compared to previous models.
Article
Full-text available
Machine learning is a means to derive artificial intelligence by discovering patterns in existing data. Here we show that applying machine learning to ordinary human language results in human-like semantic biases. We replicate a spectrum of known biases, as measured by the Implicit Association Test, using a widely used, purely statistical machine-learning model trained on a standard corpus of text from the Web. Our results indicate that text corpora contain re-coverable and accurate imprints of our historic biases, whether morally neutral as towards insects or flowers, problematic as towards race or gender, or even simply veridical, reflecting the status quo distribution of gender with respect to careers or first names. Our methods hold promise for identifying and addressing sources of bias in culture, including technology.
Article
Stereotype research emphasizes systematic processes over seemingly arbitrary contents, but content also may prove systematic. On the basis of stereotypes' intergroup functions, the stereotype content model hypothesizes that (a) 2 primary dimensions are competence and warmth, (b) frequent mixed clusters combine high warmth with low competence (paternalistic) or high competence with low warmth (envious), and (c) distinct emotions (pity, envy, admiration, contempt) differentiate the 4 competence-warmth combinations. Stereotypically, (d) status predicts high competence, and competition predicts low warmth. Nine varied samples rated gender, ethnicity, race, class, age, and disability out-groups. Contrary to antipathy models, 2 dimensions mattered, and many stereotypes were mixed, either pitying (low competence, high warmth subordinates) or envying (high competence, low warmth competitors). Stereotypically, status predicted competence, and competition predicted low warmth.
Article
Significance Word embeddings are a popular machine-learning method that represents each English word by a vector, such that the geometry between these vectors captures semantic relations between the corresponding words. We demonstrate that word embeddings can be used as a powerful tool to quantify historical trends and social change. As specific applications, we develop metrics based on word embeddings to characterize how gender stereotypes and attitudes toward ethnic minorities in the United States evolved during the 20th and 21st centuries starting from 1910. Our framework opens up a fruitful intersection between machine learning and quantitative social science.
Article
This paper proposes a simple and efficient approach for text classification and representation learning. Our experiments show that our fast text classifier fastText is often on par with deep learning classifiers in terms of accuracy, and many orders of magnitude faster for training and evaluation. We can train fastText on more than one billion words in less than ten minutes using a standard multicore CPU, and classify half a million sentences among 312K classes in less than a minute.
Article
Based on the "wisdom of the crowd" effect, consumer-generated online reviews are supposed to help consumers make more accurate product evaluations. However, the large amount of information in the reviews, coupled with conflicting opinions, can make it difficult for consumers to identify and consider those attributes relevant to their decision. Thus, while online product reviews are generally believed to empower consumers, we suggest that they may have "swaying" effects in that attribute preferences (i.e., the relative importance consumers place on various product attributes in product evaluation) are more heavily influenced by characteristics of the online reviews rather than by the relevance of the attributes to the consumers' decision context. We propose that three characteristics of online reviews affect the assessment of attribute preferences: (1) the amount of information about attribute-level performance, which is often unevenly distributed across attributes, (2) the degree of information conflict about attribute-level performance, and (3) the relationship between the overall numeric rating and the attribute-level performance information in the reviews. We test our hypotheses in two randomized experiments and a free simulation study. Results from the three studies show that the three review characteristics influence attribute preferences and that their effects are strong enough such that attribute preferences are influenced more by these online review characteristics than by the relevance of the attributes to the consumers' decision context. Our work, which illustrates a dark side to online reviews, has implications for both online word-of-mouth and preference construction research.