What are Your Pronouns? Examining Gender Pronoun Usage on Twitter

What are Your Pronouns? Examining Gender Pronoun
 Usage on Twitter
 JULIE JIANG, USC Information Sciences Institute, USA
 EMILY CHEN, USC Information Sciences Institute, USA
 LUCA LUCERI, USC Information Sciences Institute, USA
arXiv:2207.10894v1 [cs.SI] 22 Jul 2022

 GORAN MURIĆ, USC Information Sciences Institute, USA
 FRANCESCO PIERRI, USC Information Sciences Institute, USA and Politecnico di Milano, Italy
 HO-CHUN HERBERT CHANG, USC Information Sciences Institute, USC Annenberg School of Com-
 munication, USA
 EMILIO FERRARA, USC Information Sciences Institute, USC Annenberg School of Communication, USA
 Stating your gender pronouns, along with your name, is becoming the new norm of self-introductions at
 school, at the workplace, and online. The increasing prevalence and awareness of nonconforming gender
 identities put discussions of developing gender-inclusive language at the forefront. This work presents the first
 empirical research on gender pronoun usage on large-scale social media. Leveraging a Twitter dataset of over
 2 billion tweets collected continuously over two years, we find that the public declaration of gender pronouns
 is on the rise, with most people declaring as using she series pronouns, followed by he series pronouns, and a
 smaller but considerable amount of non-binary pronouns. From analyzing Twitter posts and sharing activities,
 we can discern users who use gender pronouns from those who do not and also distinguish users of various
 gender identities. We further illustrate the relationship between explicit forms of social network exposure
 to gender pronouns and their eventual gender pronoun adoption. This work carries crucial implications for
 gender-identity studies and initiates new research directions in gender-related fairness and inclusion, as well
 as support against online harassment and discrimination on social media.
 CCS Concepts: • Human-centered computing → Empirical studies in HCI; Empirical studies in col-
 laborative and social computing.
 Additional Key Words and Phrases: pronouns, gender pronouns, gender nonconformity, non-binary, gender-
 inclusivity, LGBTQ+, gender identity, social media

 1 INTRODUCTION
 Self-introductions nowadays are evolving from “Hi, my name is John” to “Hi, my name is John and
 my pronouns are they, them, theirs”. People are proactively asking, “What are your pronouns?”, a
 question that symbolizes “an invitation to declare, to honor, or to reject, not just a pronoun, but
 a gender identity” [6]. Encouraging gender pronoun sharing fosters gender-inclusive language,
 especially for those who feel that their gender identity does not align with what was assigned at
 birth [41]. Gender pronouns are becoming ubiquitous at schools, workplaces, and social media
 [17, 34, 84]. Establishments ranging from higher education to the workplace encourage the use of
 preferred pronouns to build more inclusive and safe spaces [17, 84]. Social media such as Twitter,
 Instagram, and LinkedIn are rolling out designated fields for users to add their preferred pronouns
 [25]. Gender identity issues have also been political: In the first few days of his presidency, President
 Biden signed an executive order to prevent gender discrimination based on gender identity or
 sexual orientation [42].

 Authors’ addresses: Julie Jiang, USC Information Sciences Institute, USA; Emily Chen, USC Information Sciences Institute,
 USA; Luca Luceri, USC Information Sciences Institute, USA; Goran Murić, USC Information Sciences Institute, USA;
 Francesco Pierri, USC Information Sciences Institute, USA and Politecnico di Milano, Italy; Ho-Chun Herbert Chang, USC
 Information Sciences Institute, USC Annenberg School of Communication, USA; Emilio Ferrara, USC Information Sciences
 Institute, USC Annenberg School of Communication, USA.

2 Jiang et al.

Fig. 1. The Twitter biography of the public figure Tim Cook shows he uses the he series pronouns, as seen on
July 10, 2022.

 The umbrella term, ‘non-binary’, refers to all those whose gender identity does not neatly fit
into the binary boxes of the stereotypical gender construct. Non-binary gender pronouns are
constantly evolving with the increasing worldwide acceptance of gender nonconformity and the
concept of gender as a spectrum [7, 36]. The Swedish gender-inclusive, third-person pronoun hen
achieved widespread usage and increasing acceptance since its inception in the Swedish vocabulary
in the early 2010s [36]. In English, the singular they is by far the most widely accepted gender-
neutral pronoun [39] and was voted the Word of the Decade by US linguists in 2020, beating other
contenders such as ‘meme’ and ‘#BlackLivesMatter’ [34]. But they is not the only gender-neutral
pronoun. Other popular picks include xe and ze [23], with no shortage of novel pronouns being
coined.1 Pronouns can also constantly shift for those who are gender fluid, moving across the
spectrum of gender identity over time [24, 45].
 The issue of gender identity is not without controversy [10, 12, 66, 74]. Pew Research Center
found that 54% of the US population believes that gender identity is determined by sex assigned at
birth [54, 66]. The acceptance of non-binary gender identities is tied closely with the acceptance of
LGBTQ+ people and gender dysphoria [66]. However, the share of the population who publicly use
gender-neutral pronouns or are transgenders is undeniably on the rise. According to another study
by Pew Research Center in 2019, one-in-five US adults know someone who uses gender-neutral
pronouns [32]. The proportion of people who personally know someone as transgender grew from
37% in 2017 to 42% in 2021, and the proportion of people who personally know someone who uses
gender-neutral pronouns grew from 18% in 2017 to 26% in 2021 [32]. As of 2022, an estimated 5% of
the US young adult population identifies as transgender or gender nonconforming [11].
 This paper considers the expression of gender pronouns on Twitter using a dataset collected over
the course of two years (2020-2021) on the topic of COVID-19 [16]. This dataset is selected because
it is one of the largest recent longitudinal social media datasets. More importantly, it pertains to
a topic that is not particularly relevant to gender identity, which allows us to have access to a
large sample of users who may or may not use gender pronouns. While Twitter had not formally
introduced personal pronoun fields when this data was collected, many were already including
gender pronouns in their personal biographies or display names. For instance, Tim Cook, the CEO
of Apple, included his gender pronouns in his biography (Fig 1).

1.1 Contributions
In this paper, we take an empirical and analytical approach to provide an initial understanding of
gender pronoun usage on social media, specifically to analyze the overall trends and characteristics
of gender pronoun usage and adoption at scale. Our main contributions are as follows:
 • Though users who use gender pronouns remain a minority at less than 8% at any given time,
 there is an upward trend in the number of users who openly declare gender pronouns, up
1 https://en.wikipedia.org/wiki/Gender_neutrality_in_languages_with_gendered_third-person_pronouns#Table_of_

standard_and_non-standard_third-person_singular_pronouns

What are Your Pronouns? 3

 33% from 2020 to 2021. Twice as many tweets are posted by users who identify as female
 (6 million) as are users who identify as male (3 million). A notable 1.6 million tweets are by
 users who identify as non-binary.
 • A user’s gender pronoun declaration, or lack thereof, can be distinguished from their Twitter
 usage and sharing activities.
 • Users who adopt gender pronouns experience higher social network effects of gender pronoun
 usage than those who do not, and this holds especially true for non-binary users.
With this work we hope to inform and inspire new research directions in fairness, inclusion, and
gender studies on social media, including discussions towards actionable support against online
harassment and beyond.

2 BACKGROUND AND RELATED WORK
2.1 Gender Identity and Pronouns
Gender is a deeply rooted societal concept that many believe to be synonymous with sex. But
while sex is a biological state, gender is an identity that can be either male or female, neither
male nor female, somewhere in between, or something else altogether [58]. The common social
understanding of gender maps directly from sex and dictates how people of each sex should manifest
in terms of behavior, personality, psychology, language use, capabilities, physical abilities, and
intellectual capacities, conventionally landing either in masculinity or femininity [22, 33, 58, 86, 88].
The existence of transgenders threatens this notion of well-defined boundaries of sex and gender
[75], defying gender norms deeply reinforced from popular culture to everyday discourse [15, 31].
 One of the most devastating yet overlooked social phenomenon is the high rates of suicide and
depression among transgenders and the rest of the LGBTQ+ community [14, 38, 72, 79], aggravated
by targeted hate crimes and systematic violence against gender-nonconforming people [3, 37, 43].
The anti-LGBTQ+ hate crime at a club in Florida in 2016 that claimed the lives of 49 people remains
one of the deadliest mass shootings in US history [3].
 Recognizing gender pronouns beyond the binary categories is one of the best paths forward to
developing gender-inclusive language [13, 41, 46, 62, 67]. Using a person’s correct pronouns shows
respect and validation for their chosen gender identity and self-representation [46, 67]. However,
despite the best of intentions, Levin [53] argues that mandatory gender pronoun sharing may
instigate anxiety for those who already struggling with their gender identities. That being said,
inviting—but not requiring—people to share their pronouns can help avoid predisposed assumptions
of gender identity and misgenderings [30, 63]. This simple invitation can open up conversations
about gender identity and encourage the acceptance of gender nonconformity.

2.2 Gender Pronouns on Social Media
There is a dearth of literature surrounding the use of gender pronouns on social media. Most
contributions pertaining to gender identity on social media involve developing algorithms to
classify gender or evaluate gender differences, treating gender as a binary variable [2, 5, 20, 55, 89].
This is harmful in and of itself for the gender nonconforming people we address in this work,
perpetuating biases rooted in misgendering based on sexism and cisgenderism [4]. A 2021 study by
Fosch-Villaronga et al. [28] found that Twitter’s automatic, binary gender classifier misgendered
19% of the people surveyed, including both cisgenders and non-binary people. Other studies focus
exclusively on transgenders or the LGBTQ+ [44, 50, 81], neglecting the wider cisgender society
and those who have not openly declared their gender nonconformity.
 Most similar to our work are two gender-related studies related to COVID-19 discourse on
Twitter [1, 76]. Thelwall et al. [76] conducted thematic analyses of the Twitter biography (profile)

4 Jiang et al.

descriptions of male, female, and non-binary people who tweeted about COVID-19 in the UK,
reflecting variations in dimensions such as interests, jobs, relationships, sexuality, and linguistic
styles. Al-Rawi et al. [1] explored differences in gendered COVID-19 discourse co-occurring with
men, women, and non-binary keywords, suggesting that women disproportionately experienced
domestic violence during the lockdown and non-binary people expressed concerns about blood
donation. Building on past literature, we carry out a large-scale analysis of individual users who do
and do not use gender pronouns and analyze their Twitter activities beyond biography descriptions
and text.

3 RESEARCH QUESTIONS
This work aims to fill the gaps in research in understanding the landscape of gender pronoun usage
on social media. We first examine the overall and temporal prevalence of gender pronoun usage on
Twitter based on users’ declared gender identity (or lack thereof) in their biography description.
As such, we present the preliminary research question:
RQ1: What is the prevalence of gender pronoun usage in user biographies on Twitter over time?
 Since gender pronoun declaration is an elective piece of private information that users volunteer
to supply, we question whether there are factors that could separate users with gender pronouns
from users without gender pronouns. To take this one step further, we additionally explore if we
can quantify differences among those who identify as female, male, and non-binary based on their
Twitter use. As seen in prior work, users of various gender identity groups express themselves
differently on Twitter [76]. However, the extent to which these goals can be achieved remains
unanswered. To address this gap, we combine different aspects of tweet text and Twitter activity to
answer the following question:
RQ2: Can we predict the presence of self-disclosed gender pronouns in a user’s description from
 other features?
 Finally, given the unique opportunities afforded by the longitudinal dataset we study, we explore
the mechanisms behind gender pronoun adoption on Twitter. We thus investigate whether there
are cues from a user’s Twitter usage and effects from their social networks that could help us
identify who will adopt gender pronouns. Specifically, we pose the final research question:
RQ3: Can we predict gender pronoun adoption given prior exposure and social network effects?

3.1 Definition of Terms and Phrases
To maintain consistency throughout this paper, we define the following terms and phrases. Irre-
spective of the sex assigned at birth or the gender perceived by sociocultural standards, we employ
the following definitions:
 • Female: a person who uses she series pronouns
 • Male: a person who uses he series pronouns
 • Non-binary: a person who uses gender-nonconforming pronouns (e.g., they), or a mix of
 pronouns (e.g., he/she), or is otherwise gender fluid (e.g., transitions from using she to they)
We further categorize users at a given point in time as:
 • Gender pronoun users or pronoun users: users who declare the usage of any of the gender
 pronouns in our list (see §4.1 below) either in their biographies or display names.
 • No-pronoun users: users who did not declare pronouns in their biographies or display
 names.

What are Your Pronouns? 5

Fig. 2. Left: Daily proportions of tweets (15-day rolling average) by users with gender pronoun in their
biographies or display names, displaying a statistically significant rising trend in gender pronoun usage
(Mann-Kendall Trend test, < 0.001). Right: the total number of tweets by users categorized as using she
series, he series, and non-binary pronouns in our dataset.

4 DATA
We use a public collection of nearly two years of COVID-19 Twitter data by Chen et al. [16] (release
v2.72)2 from January 21, 2020 to November 5, 2021. The dataset was collected using keywords
related to the COVID-19 pandemic, with overall 2.022 billion tweets. For each tweet, we have the
text of the tweet, the type of the tweet—original tweet, retweet, quoted tweet (retweeted tweet
with comments), or replies—and user-level information including their screen name, biography,
friends count, followers count, favourites count, statuses count, listed count, and whether they are
verified. For the retweets and quoted tweets, we also have the screen name and biography of the
user being retweeted or quoted. The data was collected in real-time, therefore we are able to track
changes in a user’s gender pronoun status over time.

4.1 Gender Pronoun Categorization
We build a regular expression to match gender pronouns that appeared in users’ biography de-
scriptions or display names. The reasons we chose to use biographical descriptions and display
names as the source of ground truth are two-fold. First, this stems from anecdotal evidence of
gender pronoun usage on Twitter. Second, the fact that using tweet text may yield false positives
since pronouns are parts of speech frequently used in normal discourse. To be matched as a gender
pronoun user, the user’s biography must contain a substring with more than one gender pronoun
separated by either forward slashes or commas, with or without additional blank spaces. We limit
our scope to three categories of gender pronouns, which are the he series pronouns: {‘he’, ‘him’,
‘his’}, she series pronouns: {‘she’, ‘her’, ‘hers’}, and gender-nonconforming pronouns: {‘they’, ‘them’,
‘theirs’, ‘their’, ‘xe’, ‘xem’, ‘ze’, ‘zem’}. The gender-nonconforming pronouns are chosen due to their
relatively widespread usage in society and in previous literature [39]. If only he series pronouns or
only she series pronouns are detected, then we label that user with the gender they identify with. If
they used gender-neutral pronouns, or a mix of gender pronouns from more than one category
(e.g., ‘he/she’, ‘she/they’), then we label them as non-binary.

5 RQ1: GENDER PRONOUN USAGE ON TWITTER
5.1 Overall Trends
We begin by taking a first look at the overall gender pronoun usage on Twitter. Fig. 2 (left) charts
the temporal patterns of the daily proportion of tweets by gender pronouns users over almost two
2 https://github.com/echen102/COVID-19-TweetIDs/releases/tag/v2.72

6 Jiang et al.

Fig. 3. Top tokens that appeared in the biography of users who use the he series pronouns, she series pronouns,
and non-binary pronouns, as well as users who do not have gender pronouns, ranked by weighted log odds.

years. Tweets by gender pronoun users represent a minority of all users’ tweets. The mean and
median of the proportion of tweets with any gender pronouns are both 3.3%, and the maximum is
7.3%. Overall, we observe a rising proportion of tweets by pronoun users. Using a Mann-Kendall
trend test, we find this growth to be statistically significant. Moreover, the average proportion
of tweets by gender pronoun users grew statistically significantly by 33% from 2.86% in 2020 to
3.82% in 2021 ( -test, < 0.001). We also see a consistent pattern of she series users (5.9 million)
dominating the gender pronoun user pool, followed by he series (3 million) and non-binary users
(1.6 million), as illustrated in Fig. 2 (right). The fact that there are twice as many tweets by females
as males stands in stark contrast with third-party Twitter surveys collected in 2019, which showed
that Twitter users consist of 50% women and 50% men — although the same research also showed
that top tweeters are more likely to be women than men [87]. Given the improbability that there
are twice as many females as males in this Twitter dataset, our work suggests that females are
more prone to declare their gender pronouns than males.

5.2 Biography Descriptors
Since we infer pronoun statuses and gender identities from biography descriptions, we assume that
there are other lexical signals contained in them. To explore that, we collect the frequencies of all
the tokens used in the biographies of she series, he series, non-binary, and no-pronoun users. We
rank the tokens using the weighted log-odds method with uniform Dirichlet priors [60], comparing
each set of tokens with the rest. We remove gender pronouns from the list of tokens since they
appear in all of the pronoun users’ biographies by definition.
 Fig. 3 shows the top tokens that appear in users’ biographies. LGBTQ+ keywords are among the
top tokens in gender pronoun users’ biographies, with ‘gay’ appearing in the biographies of he series
users and ‘queer’, ‘nonbinary’, ‘trans’, ‘lesbian’, ‘bi’, and ‘nb’ (short for ‘non-binary’) appearing in
those of non-binary users. Both she and he series users frequently use political keywords such as
‘blacklivesmatter’ (or ‘blm’ for short) and ‘feminist’. Additionally, gender pronoun users use more
interpersonal descriptors such as ‘mom’, ‘wife’, ‘auntie’, ‘dad’, and ‘husband’. Non-binary users
also use terms that suggest disability (‘disabled’, ‘autistic’). Our results echo those of Thelwall et al.
[76], which showed that gender pronoun users frequently include descriptions of interpersonal
and sexual relationships, reference politics; women and non-binary people are also more likely
to state illness and disabilities. In contrast, users without gender pronouns favor generic Twitter

What are Your Pronouns? 7

terms such as ‘news’, ‘breaking’, and ‘updates’, as well as conservative politics-related keywords.
We note that this may be skewed by Twitter accounts of official organizations or business entities,
who most likely will not include gender pronouns being that they are not personal accounts. This
requires further investigation outside the scope of this work.

6 RQ2: PREDICTING GENDER PRONOUN SELF-DISCLOSURE
In this section, we consider the extent to which we can identify a user’s gender pronoun status from
their Twitter account. Our aim is two-fold. First, we want to identify users who use any gender
pronouns from users who do not use gender pronouns. Second, we want to distinguish users of
various gender identity: male, female, or non-binary.

6.1 Method
6.1.1 Feature Sets. We use the full span of the collected data, but we consider only users who had
at least 10 posts to ensure we have sufficient data points for each user. For each user, we build a set
of features split into the following categories:
 • Text (dim = 768 × 20): the BERTweet embeddings [64] of their most recent 20 tweets, zero-
 padded if they had less than 20 tweets. We use 20 since most (60%) of the users had more
 than 20 tweets. We use BERTweet given it is the state-of-the-art transformer-based language
 model developed specifically for Twitter. Each BERT-embedding is 768 units in dimension.
 • Text Pronouns (dim = 7): the number of times I, you, we, he, she, it, and non-binary series
 pronouns (chiefly, they) appeared in their texts. Please refer to the the Appendix for the full
 list terms we used.
 • User (dim = 6): The last/latest user-level metadata including whether they are verified,
 friends count, followers count, favourites count, statuses count, and listed count.
 • Activities (dim = 9): The number of original tweets, retweets, quoted tweets, or replies the
 user posted. We also collect the number of hashtags, URLs, and mentions used, as well as the
 total number of characters and tokens used.
 • Mentions (dim = 1, 000): We first collect the top 1,000 most frequently mentioned users in
 our entire dataset. Note that in this work, we treat retweets as a type of mention. We then
 build a 1,000 dimension array where each element represents the number of times the user
 mentioned one of the top most mentioned users.
We omit users’ biographies and display names as they are our source of ground truth labeling of
gender pronouns and contain obvious cues even if we remove the gender pronoun substrings, as
seen in §5.2.
6.1.2 Modeling. We experiment with two prediction tasks. Experiment (1) is a binary classification
task of predicting whether a user displays gender pronouns or not, and experiment (2) is a multiclass
classification task of predicting the type of gender pronoun a user identifies with. To overcome data
imbalance, we sample 10,000 users from each of the gender identity groups (he series, she series,
and non-binary) followed by 30,000 users without pronouns, totaling 60,000 users. We gather two
such random samples of 60,000 users, one for hyperparameter-tuning and one for final testing via
5-fold cross-validation. For experiment (2), the 30,000 users without pronouns are not included. As
we will show below, this decision is in part driven by the relative ease in which the models were
able distinguish users with from users without pronouns. In this way, we can investigate the factors
that separate users of each gender identity group.
 We train and test a deep neural network for each prediction task. We explore using each feature set
in isolation as ablation studies, as well as all feature sets in combination. For every input scenario, we
fine-tune the model using randomized grid search to test over 40 combinations of hyperparameters,

8 Jiang et al.

Fig. 4. RQ2: Predictive performance in the binary classification of predicting the presence of gender pronouns
(left, Experiment 1) and in the multiclass classification of the three gender pronoun types (right, Experiment
2). Error bars reflect standard errors.

fixing 20% of the data as validation data. The exact model architecture hyperparameters searched
can be found in the Appendix. We use macro-F1 score as the classification metric. Since the data is
balanced, macro-F1 is equivalent to micro-F1. We use two baseline methods for comparison: Random
and Majority. The Random model predicts random labels based on the distribution of labels in the
training set and the Majority model predicts the majority label of the training set.

6.2 Results
6.2.1 Predictive Performance. We illustrate the cross-validated macro-F1 scores in Fig. 4 for experi-
ments (1) and (2). For experiment (1), the binary classification predicting gender pronoun users, we
find that all feature sets improve predictive performance over the baseline Random and Majority
models. Embeddings of the textual data are the best single predictors of gender pronouns, achieving
a macro-F1 of 84% ( = 0.05%). User activities are the second best predictors at 77% ( = 4.53%)
macro-F1, albeit with a much larger variance. Including all features yield a slightly better score of
85% ( = 0.24%). These results indicate substantial signals encoded in a user’s Twitter usage to
predict whether they disclosed gender pronouns.
 For experiment (2), the multiclass classification of which specific gender pronoun a user identifies
with, we note a similar improvement in predictive performance over the baseline. However, the
macro-F1 scores are substantially lower than what was achieved in experiment (1), with all feature
groups attaining a macro-F1 of 47% ( = 0.19%), indicating challenges in discerning users of
various gender groups. Textual features remain the best predictors of gender category at 46% macro-
F1 ( = 0.19%), whereas activities features are the worst predictor at 35% macro-F1 ( = 2.50%).
We include a breakdown of the F1 scores of each class label in the Appendix.
6.2.2 Tweet Text and Twitter Activity. We now take a closer look at the features used in the prediction
task. Fig. 5 highlights the average values of select numeric features of she series, he series, and
non-binary pronoun users, as well as users without pronouns. These features were selected to be
shown due to the substantial differences we observed. We conduct the Kruskal-Wallis H test to
ensure that they are all statistically significantly different among the four groups of users as well as
between users with and without gender pronouns (both < 0.001). A few patterns stand out. For
instance, gender pronoun users tweet considerably less than no-pronoun users. They include more
URLs but fewer hashtags. In terms of popularity metrics, gender pronouns users have fewer friends
(or followings), followers, statuses, and listed counts. However, their posts are more well-liked
(favourites counts). Gender pronoun users also use more first-person singular pronouns (‘I’) and

What are Your Pronouns? 9

Fig. 5. Average values of select numerical features for users with she, he, and non-binary pronouns as well as
users without pronouns. The number of URLs, hashtags, ‘I’ pronouns, and ‘We’ pronouns are normalized by
the number of tweets. Error bars indicate 95% confidence interval.

 She series He series Non-binary No pronouns
 URL URL URL URL 
 twitter.com 369 twitter.com 557 twitter.com 525 bit.ly 239
 mobile.twitter.com 234 nytimes.com 115 gofundme.com 258 dlvr.it 230
 paypal.me 171 washingtonpost.com 80 give2asia.org 74 ow.ly 172
 nytimes.com 147 cdc.gov 71 khalsaaid.org 68 ift.tt 159
 gofundme.com 109 tagesschau.de 70 ashaninka.fund 67 trib.al 138
 washingtonpost.com 107 cbc.ca 64 curiouscat.qa 67 foxnews.com 114
 cbc.ca 107 theatlantic.com 59 ko-fi.com 65 youtu.be 110
 soompi.com 106 npr.org 54 paypal.me 53 reut.rs 96
 theatlantic.com 85 spiegel.de 51 fundraisers.giveindia.org 53 zerohedge.com 96
 latimes.com 84 theverge.com 49 redhouseonmississippi.com 50 buff.ly 90
Table 1. The top URLs used by users with various pronoun status as determined by weighted log odds ( ).

fewer first-person plural pronouns (‘We’), displaying a higher sense of individualism as opposed to
collectivism [82].
 Since gender pronoun users use substantially more URLs in their tweets, we examine the top
URLs used by users of each gender identity group and users with no pronouns using weighted log-
odds [60]. The results (Table 1) reveal that gender pronoun users frequently cite crowdfunding or
fundraising sites such as paypal.me, gofundme.com, and fundraisers.giveindia.org, either indicating
higher interests in charitable organizations or financial distress. She and he series users also
mention more left-leaning news sites such as the New York Times, Washington Post, and the
Atlantic, whereas no-pronouns users mention more right-leaning (e.g., Fox News, Zero Hedge)
or center (e.g., Reuters) news sites.3 An explanation of this divide could be that attitudes toward
transgenders and gender nonconformity are separated along partisan lines [10, 12].
6.2.3 Highly Mentioned Users. We further explore whether there is a relationship between the
gender pronoun users and the gender identities of the 1,000 highly mentioned users. For each of
the 1,000 highly mentioned users, we check if they belong to any gender identity group. 12% of
3 Media bias ratings: https://mediabiasfactcheck.com/

10 Jiang et al.

Fig. 6. The relative frequency of a user being mentioned by a gender pronoun user as opposed to a no-
pronoun user, limited to the overall top 1,000 highly mentioned users by all users. Highly mentioned users
with pronouns, especially with non-binary pronouns, are statistically significantly more likely to be mentioned
by users also with gender pronouns.

the highly mentioned users have gender pronouns (she series: 6%, he series: 4%, and non-binary:
2%). We then compute the relative frequency of them getting mentioned by gender pronoun users
as opposed to no-pronoun users. This result is depicted in Fig 6. Across the board, all highly
mentioned users with gender pronouns are mentioned significantly more frequently by users also
with gender pronouns (Mann-Whitney U test, < 0.001). These differences are also significant
across all four groups of users using a Kruskal-Wallis H test ( < 0.001). Highly mentioned users
who are non-binary, in particular, are 10 times significantly more likely to be mentioned by gender
pronoun users than by no-pronoun users.

7 RQ3: MODELING THE ADOPTION OF GENDER PRONOUNS
As shown in §5, the overall proportion of tweets with gender pronouns is rising. As such, this section
aims to better understand the mechanisms underpinning gender pronoun adoption. Specifically,
we consider how a user’s Twitter usage and the social network they are situated in can explain
their decision to adopt or not adopt gender pronouns.

7.1 Method
7.1.1 Network Effects on Gender Pronouns. In any social network, either online or offline, people
can exert influence on others both cognitively and behaviorally, a process known as social contagion
[19, 26, 27, 40, 49, 61, 65, 71, 80]. Works have also been shown that people who congregate in social
communities share similar traits, interests, and tendencies, resulting in homophilic communities
[8, 48, 59]. It is commonly believed that processes on social networks are in actuality due to a
combination of both factors [73]. Putting network homophily and social contagion together results
in the social network effect [85]. In this work, we elicit social network effects from observable acts
of retweeting and mentioning. While it is preferable to use the follower/following relationships
among users [26, 40], this is not computationally feasible in practice given the rate limits set by the
Twitter API with a dataset of this size.4 In lieu of the following network, we use the retweet and
mention networks under the assumption that the act of retweeting or mentioning can be treated as
a form of stronger, more explicit signals of social network exposure to peer influence.
7.1.2 Defining Measurement Periods. Consider two windows time periods of data 1 and 2 , where
 1 occurred prior to 2 . We build a directed, weighted, and attributed graph = ( , ) from 1 as
follows. We first collect a set of gender pronouns users who posted tweets, were retweeted, or
4 https://developer.twitter.com/en/docs/twitter-api/rate-limits#v2-limits

What are Your Pronouns? 11

Fig. 7. The workflow used to predict and analyze how prior Twitter usage and sharing activities inform future
gender pronoun adoption.

were mentioned during 1 . Then we collect the network interactions = { }, where each edge
 = ( , , ) is a directed edge with weight representing either retweets or mentions and least
one of the two users in each edge must have gender pronouns during 1 , i.e., ∈ or ∈ .
The weight of the edge represents the frequency of the mention or retweet interactions. In the
context of this paper, retweeting a user counts as a mention, therefore the retweet graph edges are
a strict subset of the mention graph edges. We then retain the largest weakly connected component
of the graph. The users who are in the graph but did not have pronouns will be in the set ,
where = ⊔ . In the ensuing time period 2 , we check if any of the users in adopted
gender pronouns. Note that a substantial amount of users would have unknowable pronoun status
if they did not appear in 2 .
 We repeatedly build 1 and 2 data using sliding windows of discretized time periods starting
from February 1, 2020. We experiment with time spans of four weeks for graphs built on both
retweet edges and mention edges. This produces twenty-one 1 and 2 pairs and ending the last 2
period on October 9, 2021. For graphs with retweet edges, we further explore a larger time span of
eight weeks to control for data biases due to time span selection, producing ten 1 and 2 pairs and
ending the last 2 period on October 23, 2021. An illustration of this workflow is depicted in Fig. 7.
7.1.3 Features. We explore the relationship between data collected from 1 and the adoption of
gender pronoun in 2 of all users in . We use the following categories of features:
 • Text (dim = 768): the BERTweet [64] embeddings of any tweets they posted during 1 ,
 concatenated as a single string with split by periods.
 • User (dim = 768 + 6): the BERTweet [64] embedding of their biography and user-level
 metadata including whether they are verified, friends count, followers count, favourites
 count, statuses count, and listed count. We use the last/latest entry in 1 if a user appeared
 more than once.
 • Network (dim = 5 + 5 + 128): the number of neighbors they have (which is the number
 of people they retweeted from or mentioned), the number of neighbors with any gender
 pronouns, and the number of neighbors with she series, he series, or non-binary pronouns.
 We further compute all weighted versions of these metrics. Additionally, we include the
 node2vec [35] graph embeddings learned from the network (see Appendix for details).
 • Activities (dim = 11): the number of original tweets, retweets, quoted tweets, or replies
 the user posted during 1 . We also count the number of hashtags, URLs, and mentions used
 during this time period, as well as the total number of characters and tokens used. Finally,
 we include the number of first-person singular pronouns, which indicate individualism, and
 the number of first-person plural pronouns, which indicate collectivism [82].

12 Jiang et al.

 Network Time Span No. Users Pronouns (% of Total)
 (Wks) Total Pronouns She Series He series Non-binary
 Retweet Four 11,435,974 165,590 (1.4%) 0.9% 0.3% 0.3%
 Retweet Eight 4,106,182 101,023 (2.5%) 1.6% 0.5% 0.4%
 Mention Four 19,533,731 278,084 (1.4%) 0.9% 0.3% 0.2%
Table 2. The number of users included in the RQ3 experiments and the proportion of users who adopted
gender pronouns in 2 .

7.1.4 Modeling. Similar to §6.1.2, we conduct two prediction experiments. Experiment (1) is the
binary classification of whether the user will adopt gender pronouns in 2 and experiment (2) is
the multiclass classification of which of the three gender identities the user adopts in 2 . Users
who did not adopt any gender pronouns are excluded from experiment (2), as was done in §6.1.2.
Our experiments are repeated for all three data scenarios: retweet graphs over four-week time
spans, retweet graphs over eight-week time spans, and mention graphs over four-week time spans.
That is, each duration is either four or eight weeks. For each data scenario, we use the last
 1 – 2 period of data as the test set, the second to last 1 – 2 period of data as the validation set,
and the remainder of the dataset as the training set. We first tune the model hyperparameters via
randomized grid search using the training set. During hyperparameter tuning, we further break the
training dataset into sub-training, sub-validation, and sub-testing sets, using the last two periods of
data as the sub-validation and sub-testing sets, respectively. Please see the Appendix for the exact
model architecture and hyperparameter search space. The evaluation metrics used are macro-F1.
Following §6.1.2, we explore using each feature set in isolation as well as in combination. We also
use the Random and Majority models as baselines.

7.2 Results
7.2.1 Data Statistics. The adoption experiment data sets used in this section are described in
Table 2. We show the number of target users included in the experiments and how many of those
eventually adopted pronouns. These users are all part of , that is, users who were connected to
a gender pronoun user ( ) in 1 but who themselves did not have gender pronouns in 1 . Overall,
the proportion of gender pronoun adoption is very low at around 2%, rendering the data extremely
imbalanced. This comes as no surprise given the relatively low gender pronoun usage we have seen
so far in this paper, even though, by design, every user is connected to at least one gender pronoun
user through retweeting or mentioning them. Of all the gender pronouns adopted, she series is the
most popular pick. He series and non-binary pronoun adoptions following closely behind, with he
series pronouns slightly more frequently adopted.
7.2.2 Predictive Performance. We present our results of experiments (1) and (2) in Fig. 8. The
experimental results of the four-week and eight-week time spans of the retweet network are
combined due to the performance similarity. We observe challenges in experiment (1), the binary
classification of predicting whether a user will adopt any type of gender pronouns in 2 given
the features in 1 . The features we collected perform only slightly better than the Random and
Majority baseline on the retweet network, achieving 56% macro-F1 compared to the baseline score
50% macro-F1. The mention network prediction tasks achieve better performance at macro-F1
scores of 59%. The combination of all features yields very little, if at all, improvement gain over
using features in isolation. All singular categories of features perform relatively equivalently. User
features, which include their biography and other user-level metrics, are the most predictive,

What are Your Pronouns? 13

Fig. 8. RQ3: Predictive performance in the binary classification of predicting the presence of gender pronouns
(left, Experiment 1) and in the multiclass classification of the three gender pronoun types (right, Experiment
2). The top row shows the results of the retweet networks, combining retweet networks spanning four-weeks
and eight-weeks. The bottom row shows the results of the mention network spanning four-weeks.

while network features are some of the lowest. We note that this could be an artifact of our study
design limited by computational constraints, forsaking the larger social network of users who
are not directly connected to a gender pronoun user. Therefore, we have no knowledge as to
how many other, no-pronoun users they were connected to. The ineffectiveness of the network
features becomes even more apparent in the mention network prediction task. One explanation
is the underlying difference between retweets and mentions; while retweets are often seen as
endorsements [9], mentions can represent social conversations [51].
 The predictive performance for experiment (2), which is the multiclass classification of which of
the three gender identity a gender pronoun adoptee chooses, yields several interesting insights.
While user-level features remain the most singularly predictive set of features, we now see a large
improvement gained from using network features. This shows that network features contains
cues as to which gender pronoun a user will adopt. This process could unfold via either social
homophily or social contagion or both [73]. Social networks are mediums for online communities
of people with shared interests and of similar backgrounds, forming in homophilic clusters of users.
It can also be pathway for social influence, wherein one exerts behavioral or emotional changes
on another by virtue of exposure. While we cannot disentangle the intertwining effects of the
two forces, below we provide additional analyses on social network effects in relation to gender
pronoun adoption.
7.2.3 Relative Number of Pronoun Neighbors. To provide a deeper understanding of the link
connecting social networks and gender pronoun adoption, we measure the relationship between
gender pronoun adoption in 2 and the relative number of gender pronoun neighbors they had in 1 .
A user’s neighbor is someone they retweeted in a retweet network and someone they mentioned

14 Jiang et al.

Fig. 9. The relative number of neighbors with gender pronouns for a user who adopted she series pronouns,
he series pronouns, non-binary pronouns, or did not adopt gender pronouns in 2 , as given by Eq. 2, for the
retweet (left) and mention (right) networks. Users who adopted non-binary pronouns in 2 had an outsized
number of non-binary neighbors in 1 .

in a mention network. To this end, we calculate the relative number of gender pronoun neighbors
a gender pronoun adoptee had in 1 . We first establish the expected number of type gender
pronoun neighbors any no-pronoun user would be connected to in 1 , given by:

 1 ∑︁
 [ ] = N ( ), (1)
 | | ∈ 
 
where N ( ) denotes the number of -gendered neighbors a user ∈ had in 1 , ∈
{she, he, non-binary}. This figure includes all users in , including the ones with unknown
pronoun status in 2 . The relative number of gender pronoun neighbors for users with a given
pronoun status in 2 is thus given by:
 ,
 1 ∑︁
 ( , ) = N ( ) [ ], (2)
 | | ∈ 

where ⊂ represents the set of users with pronoun status in 2 , ∈ {she, he, non-binary,
no pronouns}. If ( , ) < 1, then users who adopted pronoun status in 2 had less -gendered
pronoun neighbors in 1 . Alternatively, if ( , ) > 1, then users who adopted pronoun status in
 2 had more -gendered pronoun neighbors in 1 .
 Fig. 9 displays the relative number of neighbors ( , ) separately for the retweet and mention
graphs. We aggregate the statistics for each individual in every timeframe. The retweet graphs of
four-week and eight-week time spans are combined in this figure due to the similarity of their results.
There are several key insights. First, we observe that users who did not adopt pronouns (red bars)
retweeted or mentioned an expected number of gender pronoun users. That is, ( , no pronouns) ≈
1. This is both a sanity check and a result that shows that users who remained without gender
pronouns had network exposures that did not deviate from the norm.
 Additionally, users who adopted she and he series pronouns both had slightly more retweets
or mentions of same-gendered neighbors. Interestingly, however, users who adopted she series
pronouns had slightly less network influence from male neighbors in both the retweet and mention

What are Your Pronouns? 15

Fig. 10. The average adoption rate in 2 of any pronouns, she series pronouns, he series pronoun, and non-
binary pronouns as a function of the number of retweet graph neighbors using she series, he series, and
non-binary pronouns in 1 . Three retweets from users using non-binary pronouns yielded the highest adoption
rate.

networks, and users who adopted he series pronouns also had less mention network influence from
female neighbors.
 Last but certainly not least, we see that users who adopted any gender pronouns at all had more
non-binary neighbors. Those who adopted she and he series pronouns had 1.5 times non-binary
neighbors in the retweet network. This number is particularly sizeable for non-binary pronoun
adoptees, who retweeted over 3 times as many and mentioned over 2.5 times as many non-binary
neighbors.
7.2.4 Adoption Rates Given the Number of Pronoun Neighbors. We further explore the mechanisms
of network influence on adoption rates. For this purpose, we plot the empirical probability of
gender pronoun adoption in 2 as a function of the number of pronoun neighbors they had in
 1 (Fig. 10). This figure shows that, overall, gender pronoun adoption rates of any kind peaks at
three gender pronoun neighbors. Having more he-series neighbors actually decreases rather than
increases gender pronoun adoption, especially for she series adoptees. Among the three gender
identity groups, non-binary pronoun adoptees are the most sensitive to social influence, followed
by she series adoptees and he series adoptees. For all three gender identity groups, non-binary
neighbors lead to the most considerable gain in adoption rates, especially for those who also adopt
non-binary gender identities.

8 DISCUSSION
With an estimated 1.6% of the current US population being transgender or non-binary [11], it is
imperative we call attention to the importance of developing gender-inclusive language, which
relies on correctly using a person’s preferred gender pronouns [13, 41, 46, 62, 67, 69, 78]. This work
is a critical initial step towards understanding the evolving landscape of gender pronoun use on
social media. We discuss below the important findings and implications of this paper.
 RQ1: Gender Pronoun Usage on Twitter. Our results show that gender pronoun usage is consider-
able on Twitter, with a rising number of tweets written by users with gender pronouns over time.
The most prevalent gender identity is female, followed by male and non-binary. She series users
overwhelming dominate he series user at a ratio of 2:1, which could indicate that those who identify
as female are more prone to publicly disclose their gender pronouns than those who identify as
male. While there are less non-binary gender pronoun users, they nonetheless occupy a sizable
proportion of the total Twitter population.

16 Jiang et al.

 Since we find millions of she and he series users, it would be interesting to see how many of them
are cisgenders versus transgenders. According to our definition, transgenders who identify with
she or he series pronouns would be categorized as female or male, respectively. In this work, we
deliberately stray away from automatic inferences of user gender using features that were used in
prior works—such as first names [55] or lexical features [5]—for they may perpetuate precisely the
gender biases we wish to avoid. Nevertheless, since the top tokens appearing in the biographies of
she and he series users do not relate to transgenders, we have reasons to hypothesize that most of
these gender pronoun users are cisgenders as opposed to transgenders. This would require further
analysis with human annotations to avoid gender biases from models and datasets, which we leave
for future work.
 RQ2: Predicting Gender Pronoun Self-Disclosure. We show that there are cues from a user’s Twitter
usage and sharing activities that can inform whether a user publicly declares gender pronouns,
and to which gender identity they identify with. The implications of this RQ2 are two-fold. On one
hand, with textual data being the best predictor of gender pronoun status or lack thereof, there
could be a difference in discourse revolving COVID-19 among male, female, non-binary people,
and users without gender pronouns. While previous Twitter work indeed showed that each gender
identity group grapples with different COVID-19 concerns [1], we further provide evidence that
discourse also differs between users with explicit gender identity and those who do not. On the
other hand, non-textual features can also inform predictive modeling, suggesting that there are
distinguishing patterns of social media usage that could help us discern users’ proclivity to gender
identity declaration and users of various gender identity groups. We believe this matter merits
further research efforts and we hope this work sheds light on new research directions for the
broader CSCW community.
 RQ3: Modeling Gender Pronoun Adoption. The social contagion theory stipulates that emotive
states and behaviors can spread through social networks, not unlike the transmission of pathogens
[19, 57], either intentionally or unintentionally. A wealth of literature has demonstrated that effects
of social influence on the adoption of behavior–smoking [18], eating [65], happiness [26, 29],
depression [71], generosity [68, 80], and more. With careful experiment setups, social contagion
can also be readily observed on social media [21, 26, 40, 49, 61]. However, as Shalizi and Thomas
[73] explains, extrapolating conclusions of social contagion from social network data is challenging
because “homophily and contagion are generically confounded”. Without knowing all aspects
of social influence and latent covariates of an individual–which is virtually impossible to come
by–one cannot distinguish homophily, which is the tendency of people sharing similar traits to be
connected to each other, from contagion, which is the causal impact of one individual’s actions on
another [73].
 In this work, we attempt to explain the process of gender pronoun adoption using prior explicit
signals of gender pronoun use on Twitter, operationalizing exposure and potential influence as
retweeting or mentioning another gender pronoun user. However, we recognize the limitations of
our dataset and do not posit causal relationships of social network effects we cannot infer. Gender
pronoun adoption could be explained by contagion in that exposure to other gender pronoun users
influences one’s own decision to use gender pronouns. Or it could be explained by homophily in
that users of similar gender identities naturally attracts one another.
 Ultimately, we believe that gender pronoun adoption is a much more complex process beyond
the intertwining effects of homophily and contagion. The share of transgenders and gender non-
conforming people we know personally are rising [11], more and more celebrities such as Caitlyn
Jenner and Elliot Page are coming out as transgenders. Gender pronoun usage is acquiring increas-
ing exposure at work [17] and at school [84]. We read about gender identity on the news [77] and

What are Your Pronouns? 17

discuss its repercussions on politics [52], sports [56], and law enforcement [47]. It is therefore futile
to decouple the effects of the world surrounding us on the adoption of gender pronouns on Twitter.
 With these precautions in mind, we make two conclusions regarding RQ3. The first is that
different gender identity groups responds differently to social network stimuli, possibly indicating
the each gender identity group have distinct mechanisms of susceptibility to influence. The second
conclusion is that each gender identity group exerts varying levels of influence on gender pronoun
adoption. Non-binary pronoun users appears to exert the most influence on the adoption of any
gender identity, while he and she series pronoun users exert more influence on same-gender users
but less influence on opposite-genders. One hypothesis is perhaps that males and females feel
the need to distinguish themselves from the opposite gender identity and thus deviating from the
perceived expressions of gender identity they observe. This is an interesting result that warrants
further discussion and analysis, which we leave for future work. In sum, we find that non-binary
individuals stand out as both being the most likely to adopt gender pronouns after seeing others use
gender pronouns and also being the most related to others adopting gender pronouns in subsequent
time periods.

8.1 Broader Implications
Openly declaring one’s gender nonconformity can be daunting [53]. Yet in this paper, we witness
an increasing number of people who feel comfortable enough to publicly share their non-binary
identities. The rise in gender pronoun usage, not only online but also offline, serves as an important
stepping stone towards building a society in which we acknowledge, respect, and value the spectrum
of gender identities.
 It may seem counterintuitive, but sharing pronouns is not only important for gender noncon-
forming individuals, but also for cisgenders whose gender does conform to their sex. Without
cisgenders also declaring their pronouns and normalizing the use of preferred gender pronouns,
discussions of gender pronouns equates to a deviation from gender norms, trivializing the already
stressful and painful experiences for those grappling with complex relationships with their gender
identities [69]. Encouraging everyone to use gender pronouns fosters a safe, gender-inclusive space
by signaling allyship and reaffirming the freedom to choose gender identities [41, 69, 78]. The active
declaration of gender pronouns by cisgender people, however self-evident, is an expression of “We
stand in solidarity” with those whose gender identities lie beyond the dichotomous sociocultural
perceptions of gender. Our work illustrates the effects of gender pronoun usage on gender pronoun
adoption, providing actionable insights toward improving gender inclusivity in online spaces.

8.2 Generalizability
At the time of writing (mid-2022), the COVID-19 pandemic, which began in early 2020, has been
and remains a sizable concern plaguing our everyday lives. It is the omnipresence of the pandemic
that motivated us to use the COVID-19 Twitter dataset as a tool to study a heterogeneous sample of
the Twitter population, including not only the marginalized LGBTQ+ community but also the wider
cisgender society. By focusing on more generalizable aspects of network dynamics and Twitter
usage, we believe that our results carry transferable implications to other situations, contexts, and
mediums beyond strictly COVID-19 discussions.

8.3 Ethical Consideration
While the insights we draw from this work are invaluable for gender-related studies, we underscore
the severe ethical impact of our work. Transgenders and the rest of the LGBTQ+ community
are no stranger to stigma, discrimination, human rights violation, hate speech, and hate crimes
[3, 37, 43, 83]. In many parts of the world, people with nonconforming gender identities struggle

18 Jiang et al.

with legal gender recognition [83] and suffer cataclysmic and criminalizing consequences due
to their gender identities [70]. Hence, demonstrating the possibility to predict gender pronoun
self-disclosure and adoption mechanisms on social media could lead to abuse. In this study, we
took extreme measures to avoid exposing harmful personal details, and our analyses are only in
aggregate and at a high level. But malicious actors could use the same principles to discover and
target gender nonconforming people. Therefore, we strongly encourage future studies to consider
the ethical aspects of online gender-related studies from the initial study design to the final research
output.
 Note: Both the data collection and the study have been IRB approved.

8.4 Limitations
We acknowledge several limitations of our work. First, we consider only public-facing Twitter users,
which presents unavoidable self-selection and platform biases in the users included in this study.
Second, we treat all textual data, especially gender pronouns, as English, resulting in a Western-
centric interpretation of our insights. Finally, our definition of social influence is self-selected
exposure rather than the more admissible organic, involuntary exposure. Notwithstanding these
limitations, we are optimistic that our results carry actionable insights and crucial implications
that could inform additional research in this area.

9 CONCLUSION
This paper provides a preliminary understanding of the landscape of gender pronoun usage on
Twitter in recent years. We demonstrate the rising prevalence of gender pronoun usage and illustrate
the differences between users with and without explicit gender identity, as well as between users
who identify as male, female, and non-binary. We explain how social network influence could lead
to the adoption of gender pronouns. The implications of our research offer a renewed understanding
of how non-binary people could influence and be influenced by explicit declarations of gender
pronoun usage. Our work spurs numerous exciting avenues of future research for the wider CSCW
community. In addition to those we already discussed above, one research direction would be to
better understand why some people remove their gender pronouns from their biography, perhaps
in fear of stigma or attacks. We hope our work motivates new discussions, research, and endeavors
in gender-related initiatives and human rights efforts.

10 ACKNOWLEDGEMENTS
The authors are indebted and grateful for people from the LGBTQ+ community who offered their
feedback on this work. We also thank Prof. Jonathan May (USC) for his helpful insights. This work
was supported by DARPA (award number HR001121C0169).

REFERENCES
 [1] Ahmed Al-Rawi, Karen Grepin, Xiaosu Li, Rosemary Morgan, Clare Wenham, and Julia Smith. 2021. Investigating public
 discourses around gender and COVID-19: a social media analysis of Twitter data. Journal of Healthcare Informatics
 Research 5, 3 (2021), 249–269.
 [2] Jalal S Alowibdi, Ugo A Buy, and Philip Yu. 2013. Language independent gender classification on Twitter. In Proceedings
 of the 2013 IEEE/ACM international conference on advances in social networks analysis and mining. 739–743.
 [3] Scottie Andrew. 2020. Florida observes Pulse Remembrance Day on the fourth anniversary of the Pulse nightclub
 shooting. CNN (2020). Retrieved July 13, 2022 from https://www.cnn.com/2020/06/12/us/pulse-remembrance-day-
 orlando-shooting-trnd/index.html
 [4] Y Gavriel Ansara and Peter Hegarty. 2014. Methodologies of misgendering: Recommendations for reducing cisgenderism
 in psychological research. Feminism & Psychology 24, 2 (2014), 259–270.
 [5] David Bamman, Jacob Eisenstein, and Tyler Schnoebelen. 2014. Gender identity and lexical variation in social media.
 Journal of Sociolinguistics 18, 2 (2014), 135–160.