PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract

Gender inclusivity in language has become a central topic of debate and research. Its application in the cross-lingual contexts of human and machine translation (MT), however, remains largely unexplored. Here, we discuss Gender-Neutral Translation (GNT) as a form of gender inclusivity in translation and advocate for its adoption for MT models, which have been found to perpetuate gender bias and discrimination. To this aim, we review a selection of relevant institutional guidelines for Gender-Inclusive Language (GIL) to collect and systematize useful strategies of gender neutralization. Then, we discuss GNT and its scenarios of use, devising a list of desiderata. Finally, we identify the main technical challenges to the implementation of GNT in MT. Throughout these contributions we focus on translation from English into Italian, as representative of salient linguistic transfer problems, due to the different rules for gender marking in their grammar.
From Inclusive Language to Gender-Neutral Machine Translation
Andrea Piergentili1,2, Dennis Fucci1,2, Beatrice Savoldi1,2, Luisa Bentivogli2, Matteo Negri2
1University of Trento
2Fondazione Bruno Kessler
{apiergentili,dfucci,bentivo,negri}@fbk.eu
beatrice.savoldi@unitn.it
Abstract
Gender inclusivity in language has be-
come a central topic of debate and re-
search. Its application in the cross-lingual
contexts of human and machine transla-
tion (MT), however, remains largely unex-
plored. Here, we discuss Gender-Neutral
Translation (GNT) as a form of gender in-
clusivity in translation and advocate for
its adoption for MT models, which have
been found to perpetuate gender bias and
discrimination. To this aim, we review
a selection of relevant institutional guide-
lines for Gender-Inclusive Language (GIL)
to collect and systematize useful strategies
of gender neutralization. Then, we dis-
cuss GNT and its scenarios of use, devis-
ing a list of desiderata. Finally, we identify
the main technical challenges to the im-
plementation of GNT in MT. Throughout
these contributions we focus on translation
from English into Italian, as representative
of salient linguistic transfer problems, due
to the different rules for gender marking in
their grammar.
1 Introduction
Gender bias and discrimination perpetuated
through language have become prominent top-
ics of discussion, both within and outside the
scientific community. Indeed, that language
grants the power to reproduce and reinforce
societal asymmetries has been long claimed
by feminist language critique (Lazar, 2005).
Social scientists as well have framed biased
communicative practices as a socio-political
concern, underscoring how language can reflect
the perceived value, power and status associated
with genders in society (Pauwels, 2003), e.g.,
distinguishing between Mrs./Miss for women only.
Pyscholinguists have investigated the influence
of gendered forms on cognition (Boroditsky et
al., 2003), proving how masculine generics i.e.,
masculine forms, in theory generically intended
in reference to mixed groups actually evoke a
male bias (Gygax et al., 2019). Even in the news,
gender-related issues appear ever more frequently,
e.g., to report the disciplinary consequences of
an Irish teacher who refused to use a student’s
gender-neutral pronouns.1In this context, the
demand for Gender-Inclusive Language (GIL) has
grown louder than ever. The responses, however,
have been quite disparate. Moreover, the most
longstanding debates around gender and language
have assumed a binary approach (Stahlberg et
al., 2007), thus being incompatible with the
rising recognition of non-binary gender identities
(Ackerman, 2019).
Towards the avoidance of masculine gener-
ics and the inclusion of non-binary identities,
two main linguistic strategies have been em-
ployed. The first relies on innovative linguistic el-
ements, such as neopronouns (En. ze/zir instead
of he/she/him/his/her) which are emerging from
grassroots movements. The use and the accep-
tance of such forms, however, is still highly de-
bated, and often restricted to informal communica-
tion channels like social media (Comandini, 2021).
As a viable alternative that we explore in this work,
neutral language that avoids gender marking (e.g.,
chairperson instead of chairman) has been pro-
1See https://www.cbsnews.com/amp/news/
teacher-jailed- for-contempt- of-court-
refusing-student- gender-neutral-
pronouns-ireland/.
arXiv:2301.10075v1 [cs.CL] 24 Jan 2023
posed as an inclusive solution that conforms to
standard and attested linguistic resources. Along
this line, English undoubtedly occupies a predom-
inant role, being recognized as a leader of change
towards the deployment of gender-inclusive lin-
guistic strategies (Ackerman, 2019). Compara-
tively, the situation is more complicated for other
languages, due to less timely discussions on lin-
guistic inclusivity within their speaking communi-
ties, as well as to their grammatical structures. In-
deed, grammatical gender languages such as Ital-
ian exhibit an extensive system of morphosyntactic
gender agreement, where masculine/feminine dis-
tinctions are marked for several parts of speech
(e.g., En. everybody is gone It: tutti/e sono
andati/e). As a result, neutralization strategies de-
void of gender marking can entail quite laborious
rephrasing (e.g., It. ogni persona `
eandata). In
light of the foregoing, inclusive language is in-
trinsically nuanced and multifaceted, and its im-
plementation highly language-specific. For exam-
ple, while English speakers can take advantage of a
well-accepted gender-neutral pronoun such as the
singular they,2Italian speakers cannot resort to any
similarly common devices at the moment. So, how
to move across languages in an inclusive manner?
To date, answers to this question are in need. In
fact, while some monolingual systematizations to-
wards language inclusivity have emerged (see the
analysis in Section 3), cross-lingual settings such
as translation are still largely unexplored and lack
an actionable conceptual framework of reference.
In light of recent studies highlighting that lan-
guage technologies are affected by gender bias,
the adoption of inclusive language in transla-
tion assumes a relevant role in Machine Trans-
lation (MT), too. Indeed, MT systems were
found to systematically prefer masculine forms
and stereotypical gender associations in their out-
puts, thus reinforcing bias and reiterating the
under-representation of specific groups. From
a slightly different perspective, the analysis by
Savoldi et al. (2022) actually revealed that MT
models do generate gender-neutral outputs. At
a closer inspection, however, their generation ap-
pears sporadic, unintentional, and the finding inci-
dental. The study, in fact just like other existing
work in MT (Savoldi et al., 2021) was specifi-
cally designed to explore gender bias within a bi-
2See the APA style guidelines: https://apastyle.
apa.org/style-grammar- guidelines/
grammar/singular-they
nary framework. Dedicated inquiries on gender-
neutral MT, instead, are largely absent, also hin-
dered by the current lack of data resources, bench-
marks, which in turn limit the development of in-
clusive models. Such a gap calls for dedicated
work on this topic, starting from the definition of
the objectives and the main challenges to be faced
when developing gender-neutral MT systems.
In light of this, in the present work we dis-
cuss a roadmap towards the integration of inclu-
sive language in translation, with a focus on MT.
Specifically, we discuss Gender-Neutral Trans-
lation (GNT) as an actionable form of gender-
inclusive translation; we focus on gender neu-
tralization strategies in the context of English-
Italian translation, as representative of the trans-
lation pairs which include a grammatical gender
language as target. To this aim, we present three
contributions. i) We provide a thorough analy-
sis of relevant English and Italian guidelines for
GIL, aimed to inventory established neutralization
strategies (see Section 3), in the wake of which ii)
we propose a list of desiderata for GNT (see Sec-
tion 4). Finally, iii) we discuss the challenges of
implementing such desiderata in the context of MT
(see Section 5).
2 Background
The relationship between gender and language is
socially relevant: several studies showed how the
expression of gender in language relates to the
visibility of gender groups and sexism (Wasser-
man and Weseley, 2009; Prewitt-Freilino et al.,
2012). At the same time, language mirrors social
change, as exemplified by the increase of feminine
forms in literary text, reflecting the improvement
of women’s status (Twenge et al., 2012). Such
correlations thus manifest the social and political
value of our communicative practices. Accord-
ingly, language interacts with the perception and
representation of individuals (Gygax et al., 2019;
Corbett, 2013; Stahlberg et al., 2007), and gender
expressions are socially relevant categories to talk
about the self and others. Thus, their appropriate
use is critical for both human (see Section 2.1) as
well as automatically generated (see Section 2.2)
language.
2.1 Gender and Language
The notion of gender is complex and debated, and
it has been observed and discussed extensively
through the lenses of several disciplines. Gen-
der encompasses both social and individual aspects
(Kiesling, 2019). As an overarching social catego-
rization, it relates to the status that has been tradi-
tionally associated with femininity and masculin-
ity in terms of traits, practices, or roles (Lindsey,
2015). Individually, our experience of gender con-
tributes to the construction of our identity and our
positioning within society (Crenshaw, 1991).
The notion of gender is so relevant to human ex-
perience that no language lacks expressions of fe-
maleness or maleness altogether (Stahlberg et al.,
2007). However, languages differ in how they
encode gender. For example, English is a no-
tional gender language which expresses the gender
of human referents mostly through personal pro-
nouns and possessive adjectives (e.g., he/him/hers;
she/her/hers), while sometimes also relying on
lexically gendered forms (e.g., man;woman), and
a few other elements, such as compounds (e.g.,
chairman/chairwoman). Differently, grammatical
gender languages such as Italian are characterized
by a system of morphosyntactic agreement, where
several parts of speech beside the noun (e.g., verbs,
determiners, adjectives) carry gender inflections,
as in I/Le bambini/bambine sono contenti/contente
(En. The children are happy).
Regardless of grammatical differences, gender
representation in language implicates a disparity
in the representation of the genders. It reiter-
ates and reinforces social asymmetries through dis-
criminatory practices. Part of such discrimina-
tory practices rest on normative and stereotypi-
cal principles. Androcentric normativity promotes
the masculine gender as the human prototype, the
norm, encompassing the whole human experience
(Pauwels, 2003). Women are thus treated as a gen-
dered deviation from this male default, and non-
binary experiences are excluded from representa-
tion. A typical manifestation of normativity in lan-
guage is the masculine generic, namely the use of
masculine forms as conceptually generic, neutral
(e.g., one must watch his language), when refer-
ring to mixed-gender groups or when gender is un-
known or unspecified. Stereotypes, instead, are re-
iterated and reinforced in the assumption of some-
one’s gender through associations of professional
nouns and gender (e.g., nurse = feminine, doctor =
masculine) (He, 2010), fostering gender paradigms
that limit opportunities and contribute to an unfair
social system.
In light of this, we look at Gender-Inclusive
Language3for the avoidance of stereotypical and
discriminatory language. This is a form of ver-
bal hygiene (Cameron, 1995) by which people at-
tempt to regulate language i) in conformity to cer-
tain ideals, and ii) by promoting linguistic poli-
cies that reflect them. Such policies, however,
can differ significantly, even within the context of
GIL. (Motschenbacher, 2014) mentions two rele-
vant strands of gender-related linguistic policies: i)
non-sexist, which address the structures that make
women invisible or represent them as deviant from
the male norm; and ii) non-heteronormative, which
counter language constructs that assume hetero-
sexuality as normal, desired or preferable, such
as the exclusive use of ‘binarily’ gendered forms.
Though seemingly aligned in scope, it should be
noted how these two types of policies are not al-
ways compatible. A case in point regards the non-
sexist practice of using a double gender specifica-
tion (e.g., he or she), which is binary and thus dis-
couraged in non-heteronormative language poli-
cies.
Avoiding gendered expressions in grammatical
gender languages has led to the appearance of in-
novative approaches from grassroots efforts. These
innovations, quite disparate in nature, tend to fo-
cus on direct forms of inclusive language, i.e.,
neopronouns, neomorphemes, and other resources
that allow to address and mention the referent di-
rectly, without resorting to generic terms or minc-
ing words. One of the most successful examples of
such phenomena is the personal neopronoun hen
in Swedish (Hord, 2016), which is meant to be
a gender-neutral alternative to the masculine and
feminine counterparts han and hon. In Italian, sim-
ilar innovative proposals have emerged, such as the
-u and -@neomorphemes (Comandini, 2021) which
try to retool an already existing sound of the lan-
guage in the first case and to add a new one in
the latter, in order to implement a gender-neutral
option within the Italian morphology. The adop-
tion of gender-inclusive innovations, however, is
inconsistent across different languages. While
the Swedish neopronoun hen has been accepted
3The label “inclusive language” covers a wide range of prac-
tices aimed at avoiding discrimination and denigration on any
basis (class, race, religion, and gender; see APA, 2021). Such
practices have also been given several different labels: ‘inclu-
sive’, ‘neutral’ and ‘fair’ are some of the most common ones
found in the collection of guidelines we discuss in Section 3.
To set the object of our analysis within a larger scope of inclu-
sivity, we hereby rely on the label gender-inclusive language.
by speakers and institutions, the same cannot be
said of most of its equivalents in other languages.
The Italian Accademia della Crusca, for example,
has expressed against gender-neutral innovations
to the language.4
Since gender-inclusive innovations are still con-
sidered ungrammatical and their use is discouraged
in most contexts, speakers who wish to use a GIL
tend to resort to established devices of the standard
language. The prevalent category among those
devices is that of gender-neutralization strategies,
such as the preference for epicene words, i.e.
words that are not gender-marked and can be used
regardless of the referent’s gender (e.g., spokesper-
son, as opposed to spokesman and spokeswoman).
Neutralization strategies range from simple word
choices to complex sentence formulations that,
avoiding the introduction of innovative elements
into the language, are largely acceptable. Conse-
quently, we look at gender neutralization as an ac-
tionable and acceptable form of GIL and a solid
ground for the exploration of cross-lingual scenar-
ios, namely translation and automatic translation.
2.2 Gender (Bias) and Machine Translation
Language technologies have become ubiquitous
and play a significant role in our lives, with
the danger of amplifying certain biased behaviors
(Blodgett et al., 2020). For this reason, concerns
about fairness are arising, and some of them af-
fect gender. Such technologies, indeed, have been
shown to generally discriminate against certain in-
dividuals or groups with specific gender identities
in favor of others (Sun et al., 2019, inter alia). This
tendency is defined as gender bias (Stanczak and
Augenstein, 2021).
Although affecting many monolingual tasks,
gender bias comes across even more evident in
cross-lingual scenarios, such as the case of MT,
where different languages can encode very differ-
ent gender marking mechanisms. In this task, in-
deed, models have been shown to have a tendency
towards masculine default translations (e.g., En.
The student It. Lo (M) studente) or stereotypical
translations (e.g., En. The doctors and the nurses
It. I dottori (M) e le infermiere (F)) (Cho et al.,
2019; Prates et al., 2020; Bentivogli et al., 2020;
Savoldi et al., 2021; Costa-juss`
a et al., 2020, inter
alia). Gender bias has both technical and societal
4See https://accademiadellacrusca.it/it/
consulenza/un-asterisco- sul-genere/4018
implications. From a technical perspective, the er-
roneous translation of gender clearly deteriorates
the performance of the systems. From a societal
perspective, gender bias in translation technolo-
gies may contribute to misrepresent feminine en-
tities when misgendered as male, and to reinforce
some commonplaces through stereotyped transla-
tions. At the same time, it determines an unequal
quality of service, where certain groups of users
may encounter systematic mistranslations.
Although gender bias is receiving increasing at-
tention in the research community, most efforts to
address it still do not go beyond gender binarism.
Some works have recently expressed the crucial-
ity of dealing with non-binary identities in Natural
Language Processing (NLP) (Cao and Daum´
e III,
2020; Dev et al., 2021), and two main approaches
can be found among the works that have taken this
route. Brandl, Cui and Søgaard (2022) focused
on neopronouns, showing that language models
have difficulties in processing them in Swedish
(hen), Danish (de/høn) and English (they/xe). Oth-
ers focused on standard neutral solutions, for both
text classification (Attanasio et al., 2021) and nat-
ural language generation tasks, such as gender-
neutral rewriting (Sun et al., 2021; Vanmassen-
hove et al., 2021; Attanasio et al., 2021), which
consists in converting gendered forms into their
gender-neutral counterparts (e.g., En. he/she into
singular they,chairman into chairperson). Fi-
nally, for the task of coreference resolution, Cao
and Daum´
e (2020) focused on both the solutions,
while Lauscher, Crowley and Hovy (2022) adopted
a delexicalization approach, by substituting pro-
nouns with part-of-speech placeholders. As re-
gards MT, to the best of our knowledge the sole
study addressing gender-neutral forms is by Saun-
ders, Sallis, and Byrne (2020). Working on En-
glish German/Spanish, they created a bench-
mark to assess the ability of MT systems to gener-
ate neutral target sentences. However, similarly to
Lauscher, Crowley and Hovy (2022), they avoided
choosing a specific neo-morpheme over others and
resorted to gender-neutral placeholders for articles
and inflectional morphemes (e.g., En. the trainer
Es. DEF entrenadorW\END).
Overall, adopting a neutral translation can be
seen as a path towards both the avoidance of gen-
dered inferences and the use of language that is
inclusive of different gender groups. The specific
challenges that this endeavor poses for MT, how-
ever, are not negligible. On the one hand, the
complexity of implementing neutral forms comes
from the inherent difficulties posed by grammati-
cal gender languages and by the tension between
standard neutral recommended forms and innova-
tive language-specific strategies. With this respect,
given the lack of consensus regarding the accept-
ability of the latter, Saunders, Sallis, and Byrne
(2020) refrained from picking which actual neo-
morphemes systems should be required to gener-
ate. On the other hand, the application of an in-
clusive language must be carefully designed not to
be perceived as intrusive nor as a limitation to the
freedom of expression. To exemplify this risk, a
case in point regards the assistive inclusive writ-
ing feature that has recently been incorporated in
Google Docs.5When it came out, it was strongly
rejected by several users, which unwelcomed some
seemingly spurious suggestions (e.g., the flag-
ging of motherboard as a non-inclusive term).6In
light of the foregoing, before we confront the tech-
nical challenges that arise from neutralization in
NLP, and more specifically MT, we need to lay
the groundwork for this endeavor. That is, framing
the linguistic possibilities that could be adopted to-
wards an automatic neutral translation, and identi-
fying their suitable deployment.
3 Review of Guidelines for
Gender-Inclusive Language
Looking for guidance to approach GIL, the most
influential and immediately available resources are
the guidelines produced and released by renowned
institutions to address gender discrimination in
language. We consider them ‘top-down’ ap-
proaches, as opposed to the ‘bottom-up’ efforts
of grassroots movements. Although institutional
guidelines are currently exclusively monolingual,
while our domain of interest is translation, we an-
alyze them to collect useful inclusive linguistic
strategies focusing on gender neutralization. More
precisely, we intend to i) explore how gender in-
clusivity is conceptualized within such guidelines,
and ii) gain insights concerning what should be
neutralized in language and how. To this aim, we
5See https://blog.google/products/
workspace/google-workspace- features/.
6See https://news.sky.com/story/google-
docs-criticised- for-woke- inclusive-
language-suggestions- 12598687.
selected 30 guidelines7published online8by rele-
vant institutions, equally divided between guide-
lines for English and Italian. Besides prestige,
we prioritized comparability: we selected guide-
lines by international institutions (e.g., EU) that
published the same document in both languages,
or by national institutions (e.g., universities and
governmental bodies) that share a similar status
across countries, thus also ensuring that the se-
lected guidelines belong to the same textual genre.
Starting from how these inclusive guidelines in-
terpret gender, and hence gender-based discrimi-
nation, we find clear differences between the En-
glish and the Italian documents. While the former
mostly adopt a non-heteronormative outlook ex-
plicitly going beyond the binary gender framework
the Italian guidelines tend to address women
and men only. Such a difference emerges clearly
in the two versions of the European Parliament’s
guidelines (see documents E3, I5 in the reference
list). This fundamental difference entails differ-
ent ideas of discrimination (e.g., E3: “achieving
equality”, I5: “achieving equality between men
and women”). Consequently, this conceptual dis-
crepancy is reflected in the suggested strategies to
address discrimination at the linguistic level. For
instance, the Italian guidelines provide extensive
lists of feminine counterparts for traditionally mas-
culine professional nouns (e.g., En. coordinator
as It. coordinatore [M] - coordinatrice [F]). Also,
they often endorse gender specification to avoid
masculine generics (e.g., En. The professors are
in a meeting as It. I professori [M] sono in riu-
nione -I professori [M] e le professoresse [F]
sono in riunione.) Since such suggestions remain
within a binary framework, they are irrelevant to
our gender-neutral goal, and hence discarded in
our upcoming discussion of inclusive strategies.9
Moving onto the neutralization strategies to
implement GIL, in Table 1 we offer an extensive
systematization through a multilingual perspective
that except for highly language-specific solutions
that are impossible to transfer attempts to map
strategies across English and Italian. In the ta-
ble, the strategies are presented in contrast to non-
7The full list is available at the following link
https://docs.google.com/document/d/
18vtbLQzYmTjnDsXj06rbm9FbQMHcVvpQ2Lv99xuzi2k/
edit?usp=sharing
8Retrieved through Google queries on 28/10/2022.
9Similarly, we do not treat suggestions on how to e.g., avoid
hate speech contained within these guidelines, as they exceed
the gender-neutral oriented scope of this review.
Inclusive strategy English Italian
GL Example GL Example
A. Epicene synonyms E5 Chairman
Chair,chairperson I3 Professore [Professor]
Docente [Teacher]
B. Pluralization (towards
generic or epicene forms) E2
A judge must certify that he has
familiarized himself with. . .
All judges must certify that they
have familiarized themselves with...
C. Relative and
indefinite pronouns E5
If a staff member is not satisfied...,
he can ask for a rehearing.
Any staff member who is not
satisfied... can ask for a rehearing
I3
L’assicurazione... `
e a carico
del fruitore [of the user].
a carico di chi fruisce [of
who uses]
D. Collective and Role nouns §Please contact one of the waiters.
Please contact our staff.I3
Il palazzo ospita gli studi
dei professori [of the
professors] di slavo.
. . . gli studi del personale
docente [of the teaching
staff] di slavo.
E. Omission E2
A person must reside.. . before he
may apply for permanent residence.
before applying for
permanent residence.
I3
Un’accurata compilazione...
facilita allo studente [to the
student] diverse operazioni.
facilita diverse operazioni.
F. Repetition E3
A manager may apply... if
permission has been granted by
his institution.
if permission has been granted
by that manager’s institution.
G. Passive voice E5
Each action officer must send his
document.
Documents must be sent.
I1
Il richiedente presenta la
domanda. [The applicant submits
the application]
La domanda va presentata
[The application must be
submitted].
H. Imperative forms E5
Each staff member is requested to
submit his information.
Please submit all information.
§
Il cittadino deve allegare [The
citizen must attach] un documento.
Allega [Attach] un documento.
I. Impersonal forms I15
Il candidato decade [The candidate
loses] dal diritto...
Si decade [*One loses] dal
diritto...
Table 1: Examples of neutralization strategies. In red the generic masculine formulations, in green the corresponding gender-
neutral forms. In the 2nd and 4th column, we provide the reference to the guidelines from which they were extracted (E1,2,3,..).
If no example was found for that specific strategy and language within the guidelines, but the strategy could nonetheless be
possible, we fabricated an example and indicated it with the § symbol. If a strategy is not possible to implement in one of the
two languages, the respective box is left empty.
inclusive examples. Note that their systematization
reflects different ways of realizing gender neutral-
ization, which shows higher variability.
Concerning what should be neutralized, in
fact, we identify that these documents tend to
largely focus on a particular form of gender dis-
crimination, namely, masculine generics. Mascu-
line generics have been historically employed in
administrative/legal-oriented texts to briefly refer
to the public at large (e.g., see example B, where
he refers to the whole occupational category of
judges, and the Italian il docente [M], professor
for the full teaching body). In the same vein, an-
drocentric forms are discouraged (e.g., see exam-
ple A in English). Though more briefly, it is also
discussed how GIL should avoid stereotypical as-
sociations.
This, indeed, is crucial to translation, where
cross-lingual transfer risks relying on stereotypes
(e.g., nurses rendered as F, and doctors as M in
Italian). All in all, we can see how these guide-
lines are mostly concerned with generic referents.
As we will discuss in Section 4, however, there
are also other circumstances where avoiding gen-
der marks is necessary, e.g., to avoid misgendering
individuals.
Finally, and from a linguistic standpoint, we un-
derscore that as expected English GIL strate-
gies focus on the neutralization of pronouns (e.g.,
C, E), which are the main carrier of gender distinc-
tion in notional languages. Instead, Italian guide-
lines prioritize the neutralization of nouns (see E
for an exception concerning an article), thus dis-
carding adjectives, pronouns, and verbs, which are
subject to gender agreement, too. Although the an-
alyzed sentences are simple and presented as toy
examples within an institutional genre, effective
solutions for a diverse range of communicative sit-
uations shall not overlook the full range of gen-
dered words in grammatical gender languages such
as Italian.
In light of the foregoing, we now delve into
how to avoid gender discrimination in language.
As previously anticipated, these top-down guide-
lines advocate for the use of neutralization strate-
gies that conform to standardized, institutional lan-
guage, rather than innovative, uncertain forms. As
shown in Table 1, neutral solutions can greatly
vary, ranging from omissions (e.g., E), and sim-
ple replacements of single words via unmarked
synonyms (e.g., A, B, D via epicene or collec-
tive nouns), to more complex reformulations that
involve structural changes at the sentence level
(e.g., F, G, H, I). One the one hand, though ele-
gant, nouns replacement might be limiting if other
gender-marked words are present, and only al-
low for a partial neutralization, e.g., as in It. Il
[M] professore [M] `
etenuto [M] a rispondere
(En. The professor must answer) neutralized as
L’insegnante `
etenuto [M]. When possible, how-
ever, the neutralization of short segments seems
to be a preferable dimension to consider to make
the outcome more fluent, as opposed to more com-
plex phrasings. This strategy is not always vi-
able, though. Consider, for instance, the Italian
term figlio/a (En. child): in lack of epicene syn-
onyms, neutralization would require verbose pe-
riphrases, e.g., It. minore a carico (En. under-
age, dependent child) or persona che si `
e con-
cepita o adottata (En. person who was conceived
or adopted). In light of this, neutralization strate-
gies emerge as complex choices, to be carefully
selected and weighted so as to preserve the accept-
ability of a text, i.e, fluency, style. Such choices, of
course, highly depend on various constraints (e.g.,
register, length, the context of use). Therefore,
when adopting inclusive language, it is crucial to
consider the possible trade-off between neutrality
and the overall acceptability of the text where it is
implemented.
As a final consideration, we highlight that in the
guidelines (as also reflected in Table 1) gender-
inclusive solutions are presented as a process of re-
formulation. Namely, as an additional rephrasing
step to be applied to gender-marked expressions.
To a certain extent, it resembles a (monolingual)
rewriting process, which is not ideal for transla-
tion. Indeed, the adoption of such a post-editing
approach in translation would completely exclude
the direct interpretation of the source sentence to
be neutralized, thus compromising its apt transfer
and careful decisions on when and how to neutral-
ize. With this in mind, and in light of both insights
and shortcomings that emerged in this Section, we
now specifically address the use case of Gender-
Neutral Translation.
4 Desiderata for a Gender-Inclusive
Translation
While many monolingual guidelines are avail-
able for GIL, the same cannot be said about
Gender-Inclusive Translation. Currently, results
available online for search queries like ‘gender-
inclusive translation guidelines’(En.) and ‘linee
guida traduzione inclusiva’ (It.) only provide a few
tips and tricks blog posts for translators, in addi-
tion to monolingual guidelines. This is rather un-
fortunate, as translation poses other cross-lingual
challenges. In this Section we define and dis-
cuss gender-neutral translation (GNT) as a form of
gender-inclusive translation, which allows for the
avoidance of discriminatory practices while con-
forming to the standard language. We define GNT
as a translation that does not mark the gender of
human referents if they are not assigned with a
particular gender in the source text. For exam-
ple, given the English sentence Your neighbors will
thank you, one possible Italian GNT is Il vostro
vicinato10 vi ringrazier`
a, as opposed to I vostri
10While the word vicinato is formally masculine, as a col-
lective noun referring to a group of people it is conceived as
conceptually neutral.
(1)
En.
It.
GNT
I refuse to give up on a single student in my class.
Mi rifiuto di lasciare indietro un solo studente nella mia classe.
Mi rifiuto di lasciare indietro qualsiasi studente nella mia classe.
(2)
En.
It.
GNT
A lot of innovative teachers began bringing comics...
Molti insegnanti innovativi iniziarono a portare i fumetti...
Un gran numero di insegnanti all’avanguardia iniziarono a portare i fumetti...
(3)
En.
It.
GNT
We train nurses to do it, and they use local anesthetics.
Formiamo le infermiere a farlo, e loro usano anestetici locali.
Formiamo il personale infermieristico a farlo, e loro usano anestetici locali.
(4)
En.
It.
GNT
Vehicles may only proceed at walking pace.
I veicoli possono procedere solo a passo d’uomo.
I veicoli possono procedere solo a passo di persona.
(5) En.
It.
Were you supernerdy as a girl?
Eri molto sfigata da ragazza?
(6) En.
It.
It affects one to two percent of the population, more commonly men.
Riguarda dall’uno al due percento della popolazione, ed `
e pi`
u comune negli uomini.
(7)
En.
It.
GNT
The fishermen were so upset about not having enough fish to catch that...
Ipescatori erano cos`
ıdisperati per la mancanza di pesce da pescare che...
Le persone che pescavano erano cos`
ıdisperate per la mancanza di pesce da pescare che...
(8) En.
It.
Now when I was a freshman in college, I took my first biology class.
Quando ero uno studente al primo anno di universit`
a, segu`
ı il mio primo corso di biologia.
Table 2: Examples for desiderata D1-3.
vicini vi ringrazieranno, which features a mascu-
line generic. As discussed in Section 3, the extent
to which a GNT is easy to produce, viable and ap-
propriate heavily depends on the context and the
content of the source text. Having in mind the nec-
essary trade-off between neutrality and linguistic
acceptability, as a rule of thumb we claim that
within the range of gender neutralization possibil-
ities it is preferable to choose the option that di-
verges the least from a typical and fluent transla-
tion, while maintaining semantic coherence with
the source text. Moreover, it is crucial to deter-
mine when it should be adopted. To this aim, also
informed by our analysis of the existing guidelines,
we devise three main desiderata.
D1. Avoid expressing gender in the transla-
tion when it cannot be properly assumed in the
source. While translating from one language to
another, a GNT should be used only if the gen-
der of the referent(s) cannot be inferred from the
source. This scenario could be quite frequent
in translations from a notional gender language
into a grammatical gender one. In this case, a
GNT refrains from any gratuitous assumptions,
thus avoiding expressions which may: i) misgen-
der a specific referent (see Table 2, Ex. 1); ii) ex-
clude a social group, such as in the case of mascu-
line generics (see Table 2, Ex. 2); iii) foster stereo-
typical associations (see Table 2, Ex. 3); adopt
androcentric expressions (see Table 2, Ex. 4). In
these examples, we mark in red the gender-marked
expressions falling into the binary category, and in
green those that are neutral.
D2. Use proper expressions of gender in
the translation when (indirectly) expressed in
the source. The gender of some entities can
be sometimes inferred by linguistic elements,
which we may define as “gender cues”. For
example, gender cues are 3rd person pronouns
(he/him/his,she/her/hers), terms of address (e.g.,
Mr./Mrs/Ms.), gender-specific nouns (e.g, boy,
lady, lord, wife) (see Table 2, Ex. 5). First
names, surnames, or even nicknames should not
be included among these cues for several reasons.
First, names can hardly be considered a reliable
index of someone’s gender identity (see Lauscher
et al., (2022)). Even in the attempt of any binary
correlation, names are highly ambiguous across
genders and cultures (e.g., Andrea, which is typ-
ically masculine in Italian, but feminine in Ger-
man). Clearly, the same applies to nicknames. Fi-
nally, the combination of first name and surname
is borderline: though they are attributable to a spe-
cific person, their gender expression depends on
whether that person’s gender is directly known,
and this may require a world knowledge not always
ensured (e.g., the singer Demi Lovato, who has de-
fined themselves as non-binary).11 In addition, ref-
erents’ gender could be known also through non-
textual elements, such as explicit external infor-
mation about who is speaking, which sometimes is
11See https://www.rollingstone.com/music/
music-news/demi- lovato-non- binary-
coming-out- pronouns-1171379/
provided to the translators. In all these cases, gen-
der expressions are preferable in the translation.
D3. Avoid propagating masculine generics from
source to translation. In spite of the seemingly
straightforward definition of gender cues in D2,
their recognition is not clear-cut. This is the case of
masculine generics used in the source, whose dis-
tinction from an actual gender cue might be equiv-
ocal. Hence, every information should be carefully
evaluated, in particular the compounds of -man, so
as to understand if they are used properly. For in-
stance, to explicitly refer to the masculine gender
group (see Table 2, Ex. 6), where a neutralization
would effectively compromise the meaning of the
sentence. Differently, terms such as fishermen and
freshmen are often deployed as masculine generics
to refer to the whole category (see Table 2, Ex. 7),
and thus should be translated with neutral forms.
As a final rule of thumb, however, there is a spe-
cific case where gender cues ought to be consid-
ered as largely trustworthy; namely, in relation to
the speaker as 1st person singular referent (see Ta-
ble 2, Ex. 8). This is based on the assumption
that speakers deliberately choose the most appro-
priate expressions while talking about themselves.
Hence, their choice should be respected in transla-
tion.
In conclusion, we have outlined a set of overar-
ching desiderata to guide the proposed definition
of GNT. Such desiderata have been paired with
suggestions on what and how to neutralize in trans-
lation, so as to adopt a more inclusive language
in the context of translation. Such a scaffolding
represents our proposed set of principles to be ap-
plied towards the development of more inclusive
MT models.
5 Challenges and insights for a
Gender-Neutral Machine Translation
Based on the neutralization strategies system-
atized in Section 3 and their conversion into GNT
desiderata in Section 4, we can now delve into their
applicability to MT. Indeed, on top of the theoret-
ical and linguistic insights discussed so far, the re-
search direction of a gender-neutral MT also raises
several technical challenges. These include dedi-
cated data, metrics and architectures, whose lack
necessarily hinders the development of gender-
neutral models that satisfy our desiderata.
Giving value to evaluation. The creation of
dedicated benchmarks underlies the possibility to
determine whether systems are actually making
any advancements towards automatic GNT. Ad-
hoc testbeds should frame the desired MT behavior
as outlined in the previous Section. Accordingly,
the MT benchmarks - traditionally designed as par-
allel data - should comprise source sentences that
require a GNT, aligned with their GNT counterpart
in the target language. These human translations
should be created following the strategies defined
in Section 3. Ideally, the domain of this test set
should be based on the institutional/administrative
texts from which the guidelines analyzed in Sec-
tion 3 have been derived, but other domains could
be also included to assess GNT in various con-
texts. Having this test set, however, may not be
enough: also a specific evaluation protocol needs
to be designed. Generally, automatic MT evalua-
tion is based on similarity-based comparisons be-
tween the system outputs and the corresponding
human reference translations, which represent a
single, not unique way of translating a source sen-
tence. Cast in this way, evaluation conflates and
equally penalizes i) wrong outputs and ii) accept-
able outputs that differ from the reference. Lan-
guage generation, indeed, is an open-ended chal-
lenge, for which multiple solutions are possible,
especially when periphrases are involved. For this
reason, the creation of gender-neutral reference
translations is a particularly delicate process, since
automatic neutralizations could prove to be hardly
detectable in an evaluation pipeline based on single
references. In fact, only through manual analyses
Savoldi et al. (2022) found some neutral expres-
sions in the outputs of automatic models. More-
over, standard MT evaluation metrics like BLEU
(Papineni et al., 2002) only provide a holistic score
as an indicator of overall translation quality. In-
stead, assessing GNT would require a specific pro-
tocol and measure, able to focus on the neutralized
words and isolate them from other unrelated fac-
tors that might affect generic performance. There-
fore, a possible solution could be the use of multi-
ple references, containing different neutral realiza-
tions (to account for language variability) as well
as a gender-marked realization (to identify unde-
sired outputs). Since manually creating multiple
reference translations is very costly, an alternative
solution could be designing a GNT-oriented qual-
ity estimation metric (Specia et al., 2018), which
provides an estimate of whether the output of a MT
system is gender-neutral or gender-marked, with-
out accessing any reference.
Training models without GNT examples. For
current data-driven, neural approaches to MT,
large availability of data represents a main bottle-
neck. If the manual creation and annotation of a
few hundreds of test sentences (e.g., to identify
source cues, and to determine the most appropri-
ate neutralization strategy) is a viable option for
evaluation, the vast amount of training data needed
for model development makes manual solutions
hardly feasible. To counter data lacking, it would
be interesting to explore training-time algorithmic
techniques dedicated to teaching systems how to
perform GNT. For inspiration, we could approach
gender-neutralization through the lens of lexically-
constrained MT, i.e., the task of performing a
translation in adherence to a set of pre-specified
lexical constraints (Hokamp and Liu, 2017; Post
and Vilar, 2018; Dinu et al., 2019; Bergmanis and
Pinnis, 2021, inter alia). Drawing from this line
of research, the envisioned “neutrally-constrained”
MT could rely on the creation of bilingual dictio-
naries mapping source words to desired neutral
target words to be leveraged towards the genera-
tion of epicene forms. As examined in Section 3,
however, more invasive interventions towards neu-
trality are sometimes unavoidable. Interventions
that involve more than simple word substitutions
do entail a higher cost in terms of adherence to
the source text and fluency. Hence, the ideal MT
systems should be able to consider the full range
of possible neutralization strategies, so as to plan
and find the preferable course of actions starting
from the least to the costliest ones. To this aim, a
specific training methodology that rewards highly-
probable and low-cost outputs, while penalizing
less-probable and costlier ones, could be designed
to foster minor interventions.
Disambiguating gender through a wider con-
text. Currently, state-of-the-art MT works at the
sentence level, i.e., by translating each sentence
in isolation. Although this approach still yields
the most competitive results, alternative solutions
that account for larger textual context in transla-
tion might be more apt for GNT; in particular, to
decide when to perform GNT. As discussed in Sec-
tion 4, D2-3, in fact, we should not always neutral-
ize, but rather make informed decisions depending
on the presence (or lack) of gender cues. These
cues, however, might be available out of the lim-
ited sentence-context (e.g., He was talking with a
young man. Only later I realized that this person
was his professor). Accordingly, the design of
MT models that translate beyond the sentence level
ought to be considered. Translating sentences in a
wider context, indeed, has proven crucial for cor-
rectly handling discourse cohesion (Bawden et al.,
2018), and was shown to a certain extent benefi-
cial to mitigate gender bias (Basta et al., 2020).
However, it remains occasionally dubious whether
context provides a useful linguistically-motivated
knowledge (Kim et al., 2019). Before venturing
into any document-level endeavor, it is thus recom-
mended to verify whether there is a positive inter-
pretable link between GNT, context-informed MT
which goes beyond the sentence level, and overall
quality of the system.
Disambiguating gender through external
knowledge. Another architectural choice in-
spired by D2-3 is related to the integration into the
model of external information, such as metadata
like speakers’ gender information, i.e., their
chosen linguistic expressions of gender (Gaido et
al., 2020), as a cue to disambiguate gender and
avoid unwanted neutralizations. Although MT
systems generally exploit only textual data infor-
mation from the source sentence, the injection of
external knowledge is in principle beneficial for
gender disambiguation and, in turn, for our GNT
purposes. Metadata could be supplied in the form
of tags, either at the word level (Saunders et al.,
2020; Stafanoviˇ
cs et al., 2020) or at the sentence
level (Elaraby et al., 2018; Vanmassenhove et al.,
2018; Basta et al., 2020).
6 Conclusions
Gender bias and discrimination in language are
rising concerns that have also invested the field
of automatic translation, where MT models have
been found to amplify male visibility and stereo-
types. As a promising route forward to counter
this bias, in this work we have taken the first steps
towards the adoption of gender-inclusive language
in translation. In particular, we have focused on
the use of neutral forms devoid of gender marking
for an English-Italian setting, so to avoid undue
gendering and foster the inclusion of non-binary
identities. To this aim, we have conducted an ex-
tensive review of the gender neutralization strate-
gies presented in English and Italian monolingual
guidelines for inclusivity. On this basis, we have
outlined a definition of gender-neutral translation
suitable for cross-lingual contexts, paired with a
set of overarching desiderata to guide its adoption.
Finally, we have concluded by weaving together
theoretical and linguistic insights, and technical
challenges for the implementation of a gender-
neutral translation in MT. It is in our hope that our
contribution will guide the community towards the
development of fairer and more inclusive transla-
tion practices and models.
Acknowledgements
This work has been partially funded by This work
is part of the project “Bias Mitigation and Gender
Neutralization Techniques for Automatic Transla-
tion”, which is financially supported by an Ama-
zon Research Award AWS AI grant.
References
Ackerman, Lauren. 2019. Syntactic and cognitive is-
sues in investigating gendered coreference. Glossa:
a journal of general linguistics, 4(1).
Attanasio, Giuseppe, Salvatore Greco, Moreno La Qua-
tra, Luca Cagliero, Michela Tonti, Tania Cerquitelli,
and Rachele Raus. 2021. E-mimic: Empower-
ing multilingual inclusive communication. In 2021
IEEE International Conference on Big Data (Big
Data), pages 4227–4234.
Basta, Christine, Marta R. Costa-juss`
a, and Jos´
e A. R.
Fonollosa. 2020. Towards mitigating gender bias
in a decoder-based neural machine translation model
by adding contextual information. In Proceedings
of the The Fourth Widening Natural Language Pro-
cessing Workshop, pages 99–102, Seattle, USA, July.
Association for Computational Linguistics.
Bawden, Rachel, Rico Sennrich, Alexandra Birch, and
Barry Haddow. 2018. Evaluating discourse phe-
nomena in neural machine translation. In Proceed-
ings of the 2018 Conference of the North Ameri-
can Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume
1 (Long Papers), pages 1304–1313, New Orleans,
Louisiana, June. Association for Computational Lin-
guistics.
Bentivogli, Luisa, Beatrice Savoldi, Matteo Negri, Mat-
tia A. Di Gangi, Roldano Cattoni, and Marco Turchi.
2020. Gender in Danger? Evaluating Speech Trans-
lation Technology on the MuST-SHE Corpus. In
Proceedings of the 58th Annual Meeting of the Asso-
ciation for Computational Linguistics, pages 6923–
6933, Online, July. Association for Computational
Linguistics.
Bergmanis, Toms and M¯
arcis Pinnis. 2021. Facilitat-
ing terminology translation with target lemma anno-
tations. In Proceedings of the 16th Conference of
the European Chapter of the Association for Com-
putational Linguistics: Main Volume, pages 3105–
3111, Online, April. Association for Computational
Linguistics.
Blodgett, Su Lin, Solon Barocas, Hal Daum ´
e III, and
Hanna Wallach. 2020. Language (Technology) is
Power: A Critical Survey of “Bias” in NLP. In Pro-
ceedings of the 58th Annual Meeting of the Asso-
ciation for Computational Linguistics, pages 5454–
5476, Online, July. Association for Computational
Linguistics.
Boroditsky, Lera, Lauren A. Schmidt, and Webb
Phillips. 2003. Sex, syntax, and semantics. In Get-
ner, Dedre and Susan Goldin-Meadow, editors, Lan-
guage in Mind: Advances in the Study of Language
and Thought, pages 61–79. MIT Press.
Brandl, Stephanie, Ruixiang Cui, and Anders Søgaard.
2022. How conservative are language models?
adapting to the introduction of gender-neutral pro-
nouns. In Proceedings of the 2022 Conference of
the North American Chapter of the Association for
Computational Linguistics: Human Language Tech-
nologies, pages 3624–3630, Seattle, United States,
July. Association for Computational Linguistics.
Cameron, Deborah. 1995. Verbal Hygiene (The Poli-
tics of Language. Routledge.
Cao, Yang T. and Hal Daum´
e III. 2020. Toward
Gender-Inclusive Coreference Resolution. In Pro-
ceedings of the 58th Annual Meeting of the Asso-
ciation for Computational Linguistics, pages 4568–
4595, Online, July. Association for Computational
Linguistics.
Cho, Won Ik, Ji Won Kim, Seok Min Kim, and
Nam Soo Kim. 2019. On Measuring Gender Bias
in Translation of Gender-neutral Pronouns. In Pro-
ceedings of the First Workshop on Gender Bias in
Natural Language Processing, pages 173–181, Flo-
rence, Italy, August. Association for Computational
Linguistics.
Comandini, Gloria. 2021. Salve a tutt@, tutt*, tuttu,
tuttx e tutt@: l’uso delle strategie di neutralizzazione
di genere nella comunit`
a queer online. : Indagine su
un corpus di italiano scritto informale sul web. Testo
e Senso, (23):43–64, dic.
Corbett, Greville G. 2013. The Expression of Gender.
De Gruyter.
Costa-juss`
a, Marta R., Christine Basta, and Gerard I.
G´
allego. 2020. Evaluating gender bias in speech
translation.
Crenshaw, Kimberle Williams. 1991. Mapping the
margins: Intersectionality, identity politics, and vio-
lence against women of color. Stanford Law Review,
43(6):1241–99.
Dev, Sunipa, Masoud Monajatipoor, Anaelia Ovalle,
Arjun Subramonian, Jeff Phillips, and Kai-Wei
Chang. 2021. Harms of Gender Exclusivity and
Challenges in Non-Binary Representation in Lan-
guage Technologies. In Proceedings of the 2021
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 1968–1994, Online and
Punta Cana, Dominican Republic, November. Asso-
ciation for Computational Linguistics.
Dinu, Georgiana, Prashant Mathur, Marcello Federico,
and Yaser Al-Onaizan. 2019. Training neural ma-
chine translation to apply terminology constraints.
In Proceedings of the 57th Annual Meeting of the
Association for Computational Linguistics, pages
3063–3068, Florence, Italy, July. Association for
Computational Linguistics.
Elaraby, Mostafa, Ahmed Y. Tawfik, Mahmoud Khaled,
Hany Hassan, and Aly Osama. 2018. Gender-Aware
Spoken Language Translation Applied to English-
Arabic. In Proceedings of the 2nd International
Conference on Natural Language and Speech Pro-
cessing (ICNLSP), pages 1–6, Algiers, Algeria.
Gaido, Marco, Beatrice Savoldi, Luisa Bentivogli, Mat-
teo Negri, and Marco Turchi. 2020. Breeding
Gender-aware Direct Speech Translation Systems.
In Proceedings of the 28th International Confer-
ence on Computational Linguistics, pages 3951–
3964, Online, December. International Committee
on Computational Linguistics.
Gygax, Pascal Mark, Daniel Elmiger, Sandrine Zuf-
ferey, Alan Garnham, Sabine Sczesny, Lisa von
Stockhausen, Friederike Braun, and Jane Oakhill.
2019. A Language Index of Grammatical Gender
Dimensions to Study the Impact of Grammatical
Gender on the Way We Perceive Women and Men.
Frontiers in Psychology, 10.
He, Guimei. 2010. An analysis of sexism in english.
Journal of Language Teaching and Research, 1(3).
Hokamp, Chris and Qun Liu. 2017. Lexically con-
strained decoding for sequence generation using grid
beam search. In Proceedings of the 55th Annual
Meeting of the Association for Computational Lin-
guistics (Volume 1: Long Papers), pages 1535–1546,
Vancouver, Canada, July. Association for Computa-
tional Linguistics.
Hord, Levi CR. 2016. Bucking the linguistic bi-
nary: Gender neutral language in English, Swedish,
French, and German. Western Papers in Linguis-
tics/Cahiers linguistiques de Western, 3(1):4.
Kiesling, Scott F. 2019. Language, Gender, and Sexu-
ality: An introduction. Routledge.
Kim, Yunsu, Duc Thanh Tran, and Hermann Ney.
2019. When and Why is Document-level Context
Useful in Neural Machine Translation? In Pro-
ceedings of the Fourth Workshop on Discourse in
Machine Translation (DiscoMT 2019), pages 24–
34, Hong Kong, China, November. Association for
Computational Linguistics.
Lauscher, Anne, Archie Crowley, and Dirk Hovy.
2022. Welcome to the Modern World of Pronouns:
Identity-Inclusive Natural Language Processing be-
yond Gender. In Proceedings of the 29th Inter-
national Conference on Computational Linguistics,
pages 1221–1232, Gyeongju, Republic of Korea,
October. International Committee on Computational
Linguistics.
Lazar, Michelle M., 2005. Politicizing Gender in Dis-
course: Feminist Critical Discourse Analysis as Po-
litical Perspective and Praxis, pages 1–28. Palgrave
Macmillan UK, London.
Lindsey, L.L. 2015. Gender Roles: A Sociological
Perspective. Taylor & Francis.
Motschenbacher, Heiko. 2014. Grammatical gender as
a challenge for language policy: The (im)possibility
of non-heteronormative language use in German ver-
sus English. Language policy, 13(3):243–261.
Papineni, Kishore, Salim Roukos, Todd Ward, and
Wei-Jing Zhu. 2002. BLEU: a Method for Auto-
matic Evaluation of Machine Translation. In Pro-
ceedings of the 40th Annual Meeting of the Asso-
ciation for Computational Linguistics, pages 311–
318, Philadelphia, Pennsylvania, July. Association
for Computational Linguistics.
Pauwels, Anne, 2003. Linguistic Sexism and Femi-
nist Linguistic Activism, chapter 24, pages 550–570.
John Wiley & Sons, Ltd.
Post, Matt and David Vilar. 2018. Fast lexically con-
strained decoding with dynamic beam allocation for
neural machine translation. In Proceedings of the
2018 Conference of the North American Chapter of
the Association for Computational Linguistics: Hu-
man Language Technologies, Volume 1 (Long Pa-
pers), pages 1314–1324, New Orleans, Louisiana,
June. Association for Computational Linguistics.
Prates, Marcelo O. R., Pedro H. Avelar, and Lu´
ıs C.
Lamb. 2020. Assessing gender bias in machine
translation: a case study with google translate. Neu-
ral Computing and Applications, 32(10):6363–6381,
May.
Prewitt-Freilino, Jennifer L., T. Andrew Caswell, and
Emmi K. Laakso. 2012. The gendering of language:
A comparison of gender equality in countries with
gendered, natural gender, and genderless languages.
Sex Roles, 66(3):268–281, Feb.
Saunders, Danielle and Bill Byrne. 2020. Reducing
Gender Bias in Neural Machine Translation as a Do-
main Adaptation Problem. In Proceedings of the
58th Annual Meeting of the Association for Compu-
tational Linguistics, pages 7724–7736, Online, July.
Association for Computational Linguistics.
Saunders, Danielle, Rosie Sallis, and Bill Byrne. 2020.
Neural Machine Translation Doesn’t Translate Gen-
der Coreference Right Unless You Make It. In
Proceedings of the Second Workshop on Gender
Bias in Natural Language Processing, pages 35–43,
Barcelona, Spain (Online), December. Association
for Computational Linguistics.
Savoldi, Beatrice, Marco Gaido, Luisa Bentivogli, Mat-
teo Negri, and Marco Turchi. 2021. Gender bias in
machine translation. Transactions of the Association
for Computational Linguistics, 9:845–874.
Savoldi, Beatrice, Marco Gaido, Luisa Bentivogli, Mat-
teo Negri, and Marco Turchi. 2022. Under the
Morphosyntactic Lens: A Multifaceted Evaluation
of Gender Bias in Speech Translation. ArXiv e-prints
arXiv:2203.09866. Accepted at ACL 2022.
Specia, L., C. Scarton, and G.H. Paetzold. 2018. Qual-
ity estimation for machine translation, volume 39 of
Synthesis Lectures On Human Language Technolo-
gies. Morgan & Claypool Publishers, September. ©
2018 by Morgan & Claypool.
Stafanoviˇ
cs, Art¯
urs, Toms Bergmanis, and M¯
arcis Pin-
nis. 2020. Mitigating gender bias in machine trans-
lation with target gender annotations.
Stahlberg, Dagmar, Friederike Braun, Lisa Irmen, and
Sabine Sczesny. 2007. Representation of the Sexes
in Language. Social communication, pages 163–
187.
Stanczak, Karolina and Isabelle Augenstein. 2021. A
Survey on Gender Bias in Natural Language Pro-
cessing. arXiv preprint arXiv:2112.14168.
Sun, Tony, Andrew Gaut, Shirlyn Tang, Yuxin Huang,
Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth
Belding, Kai-Wei Chang, and William Yang Wang.
2019. Mitigating Gender Bias in Natural Language
Processing: Literature Review. In Proceedings of the
57th Annual Meeting of the Association for Compu-
tational Linguistics, pages 1630–1640, Florence, IT,
July. Association for Computational Linguistics.
Sun, Tony, Kellie Webster, Apu Shah, William Yang
Wang, and Melvin Johnson. 2021. They, Them,
Theirs: Rewriting with Gender-Neutral English.
arXiv preprint arXiv:2102.06788.
Twenge, Jean M., W. Keith Campbell, and Brittany
Gentile. 2012. Male and female pronoun use in
u.s. books reflects women’s status, 1900–2008. Sex
Roles, 67(9):488–493, Nov.
Vanmassenhove, Eva, Christian Hardmeier, and Andy
Way. 2018. Getting Gender Right in Neural Ma-
chine Translation. In Proceedings of the 2018 Con-
ference on Empirical Methods in Natural Language
Processing, pages 3003–3008, Brussels, Belgium,
October-November. Association for Computational
Linguistics.
Vanmassenhove, Eva, Chris Emmery, and Dimitar
Shterionov. 2021. NeuTral Rewriter: A rule-based
and neural approach to automatic rewriting into gen-
der neutral alternatives. In Proceedings of the 2021
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 8940–8948, Online and
Punta Cana, Dominican Republic, November. Asso-
ciation for Computational Linguistics.
Wasserman, Benjamin D. and Allyson J. Weseley.
2009. ¿qu´
e? quoi? do languages with grammati-
cal gender promote sexist attitudes? Sex Roles: A
Journal of Research, 61:634–643.
... Words marked with masculine suffixes have traditionally been used in a generic sense (e.g. Madam Chairman), however, with the emergence of feminist language reform, style guides have advised against their use (Piergentili et al., 2023b). In English, the most common replacement strategy for gendered generics is neutralisation (chairperson), because all gender identities, not just male and female, can be referred to by gender-neutral nouns. ...
Article
Full-text available
Gli ultimi mesi hanno visto protagonista del dibattito politico l’uso dello schwa come morfema di genere neutro, adatto a far riferimento alle persone non binarie, che non si identificano né nel genere femminile, né in quello maschile. Questo studio propone un’analisi delle diverse strategie di neutralizzazione di genere sub-standard italiane, utilizzate nella comunicazione scritta e informale delle comunità LGBTQIA+ online,con l’obiettivo di offrire una prima panoramica preliminare sull’effettivo uso che viene fatto di questi fenomeni. Attraverso l’annotazione manuale di CoGeNSI (Corpus of Gender Neutralization Strategies in Italian), un corpus formato da testi prodotti su pagine Facebook queer, si mostreranno le regolarità e le irregolarità nell’applicazione di queste strategie, le differenze d’uso tra i riferimenti alla propria persona e quelli ad altre persone, e il loro raro uso prettamente politico. Infine, si rifletterà su come queste strategie di neutralizzazione di genere sub-standard siano un fenomeno ampiamente caratterizzato dalla velocità e dall’espressività della comunicazione mediata dal computer informale.
Article
Full-text available
Machine translation (MT) technology has facilitated our daily tasks by providing accessible shortcuts for gathering, processing, and communicating information. However, it can suffer from biases that harm users and society at large. As a relatively new field of inquiry, studies of gender bias in MT still lack cohesion. This advocates for a unified framework to ease future research. To this end, we: i) critically review current conceptualizations of bias in light of theoretical insights from related disciplines, ii) summarize previous analyses aimed at assessing gender bias in MT, iii) discuss the mitigating strategies proposed so far, and iv) point toward potential directions for future work.
Conference Paper
Preserving diversity and inclusion is becoming a compelling need in both industry and academia. The ability to use appropriate forms of writing, speaking, and gestures is not widespread even in formal communications such as public calls, public announcements, official reports, and legal documents. The improper use of linguistic expressions can foment unacceptable forms of exclusion, stereotypes as well as forms of verbal violence against minorities, including women. Furthermore, existing machine translation tools are not designed to generate inclusive content. The present paper investigates a joint effort of the research communities of linguistics and Deep Learning Natural Language Understanding in fighting against non-inclusive, prejudiced language forms. It presents a methodology aimed at tackling the improper use of language in formal communication, with a particular attention paid to Romanic languages (Italian, in particular). State-of-the-art Deep Language Modeling architectures are exploited to automatically identify non-inclusive text snippets, suggest alternative forms, and produce inclusive text rephrasing. A preliminary evaluation conducted on a benchmark dataset shows promising results, i.e., 85% accuracy in predicting inclusive/non-inclusive communications.