This study focuses on the growing importance of gender-fair language and explores innovative strategies proposed also in other languages to avoid gender-specific endings. We present a set of guidelines for the annotation and reformulation of gender-(un)fair texts and their application to a corpus of 1,024 portions of university administrative documents in Italian. Overall, the guidelines presented in this study prove to be valuable both practically and theoretically. They help identify and address non-inclusive expressions while highlighting the complexities of obscuration and visibility in gender-fair language reformulation. In addition, the statistical analysis of the created corpus shows how administrative texts tend to contain gender-unfair language, especially the masculine overextended expressions, showing the need to adopt specific and complete guidelines that lead (and support the staf training) to the use of a more gender-fair language.
Recognizing the importance of rectifying such lan- (in)visibilization of women in language, more recent
scholguage discrepancies, various guidelines have been pub- arly and activist debates also concern how to address and
lished over the years [e.g., 5]. The annotation scheme talk about non-binary people, i.e., those that do not
excluwe propose draws upon the recommendations presented sively identify as men or women, aiming at making them
in the available guidelines, to develop a comprehensive visible too.
framework for addressing gender-unfair expressions in The Italian grammatical gender system, indeed, is
biItalian language usage. To the best of our knowledge, our nary and does not provide a straightforward way to refer
annotation scheme represents a novel approach. While to non-binary people. Various gender-neutral sufixes are
another project (i.e., E-MIMIC) focuses on inclusive lan- in use, such as ‘-@’ or ‘-u’ (see [11] for a comprehensive
guage, it simply distinguishes between inclusive and non- list). As González Vázquez et al. [12] observe, such
innoinclusive texts [
In this work, firstly we review previous studies on erential gender is expressed by diferent lexical roots (e.g., this topic, both in theoretical linguistics (subsection 2.1) “madre”, mother:f.sg, and “padre”, father:m.sg); nouns and in NLP (subsection 2.2). We then describe in detail with mobile gender, whose referential gender is specthe annotation scheme (section 3) and the creation of the ified through the addition of diferent sufixes to the annotated corpus (section 4), also providing a preliminary same lexical root (e.g., “figlia”, daughter: f.sg, and “figlio”, analysis of the data gathered so far. son:m.sg); and the so-called epicene nouns, whose gender is not overtly marked, but only revealed by satellite elements – i.e., the noun’s determiners and modifiers 2. Related Work (e.g., “la nipote”, the.f.sg niece:f.sg; “il nipote”, the.m.sg nephew:m.sg). As Formato [14] observes, some nouns 2.1. Linguistics (i.e., ‘semi-epicene’) work in the latter way only in the singular and have diferent gendered sufixes in the plural (e.g., “giornalista”, journalist:f/m.sg; “giornaliste”, journalist:f.pl; “giornalisti”, journalist:m.pl). Finally, a few nouns refer to individuals of any gender irrespective of their grammatical gender (e.g., “persona”, person:f.sg).
Due to this peculiar characteristic, these nouns can be straightforwardly used to refer to non-binary people as well. Moreover, gender-neutral sufixes are not required for epicene (and, in the singular, semi-epicene) nouns, as they are not overtly marked for gender. In this case, the only needed precaution to get a gender-neutral form concerns the choice of gender-neutral satellite elements or their gender-neutralization. Gender-neutral sufixes are further inefective for nouns like madre and padre, where it is the root to be overtly marked for gender. These word endings, thus, should only be used with nouns with mobile gender, which, however, constitute the vast majority of Italian animate nouns (see [15], p. 106). Innovative strategies should also be used for the many gendered pronouns, determiners, past participles, and adjectives in order to make them gender-neutral and suitable to refer to non-binary people and to mixedgender groups.
Formato [14] also provides a taxonomy of linguistic us
are increasingly studied and debated topics (see [
Sabatini [
Her recommendations have been expanded and adapted by several private and public bodies, that issued gender-fair language guidelines [e.g., 9, 5, 10]. These works, among others, distinguish between strategies aimed at symmetrizing language by giving women the same visibility that men have, and strategies aimed at getting rid of sexism through the avoidance of gendered forms altogether. While Sabatini and the subsequent guidelines focus on the discrimination and
ages influenced by gendered assumptions. Like Sabatini, Formato focuses on both sexist expressions and linguistic conventions. Among the latter, Formato originally distinguishes the case in which masculine terms are used for mixed-gender groups from those in which they are used for unknown or generic individuals.
In our framework, we elaborate on the categories identified in these works, to develop our own taxonomy, that we implement in the annotation scheme described in section 3. 2.2. Natural Language Processing the abundance of papers on gender bias, many newly developed algorithms lack suficient bias testing and fail to address ethical considerations. Lastly, the methodologies employed in this area often lack comprehensive definitions of gender bias and robust evaluation baselines and pipelines.
The present work contributes to address many of these issues: we explicitly take into account the linguistic representation of non-binary individuals, we create an annotation scheme for Italian – for which much fewer NLP studies are available compared to English – and ifnally, we present an annotated corpus that could be exploited for the training of automatic NLP tools.
In recent years, sexism and gender-(un)fair practices have been addressed in Computational Linguistics, mostly focusing on the presence of gender bias in automatic sys- 3. Annotation Scheme tems. As highlighted in Costa-jussà [16], studies on gender bias in NLP serve a dual role. On the one hand, NLP The annotation task is divided into two parts. A first can function as a tool to identify gender bias in various so- annotation layer concerns the identification of portions cial domains such as online news or advertisements. On of text(s) where gender-unfair language is used, and the the other hand, it frequently generates gender-biased assignment of each of them to a specific type among the systems, thus contributing to the perpetuation and rein- following ones: forcement of gender bias within society. This bias in NLP • ‘incongruous’ (It. “incongruo”), when the gramis predominantly attributed to the training of models on matical gender of the noun (and, possibly, of its datasets that exhibit inherent biases. Consequently, the modifiers), does not match the gender of the refamplification of bias occurs through the learning algo- erent identified in discourse (e.g., “il ministro rithms employed in NLP systems. del turismo, Daniela Santanché”, the.m.sg min
Some specific studies have been conducted in the field ister:m.sg of.the tourism, Daniela Santanché); of Machine Translation. One of the most recent was car- • ‘overextended’ (It. “sovraesteso”), when the masried out by Rescigno et al. [17] who explored how three of culine (or, in rare cases, feminine) grammatical the most popular translation systems (Google Translate, gender is used to refer to a group composed of Bing Microsoft Translator, and DeepL) handle gender people with diferent genders (e.g., “il rapporto phenomena in natural languages, such as pronouns, job con i professori è buono”, the relationship with titles, and occupation names. The authors compared the the.m.pl professor:m.pl is good, with reference translations generated from English respectively to Ital- to a group of teachers possibly comprising men, ian, French, and Spanish, revealing that all three systems women, and non-binary individuals); exhibit some level of gender bias, with Google Trans- • ‘generic’ (It. “generico”) when the masculine (or, late producing more biased translations, Bing Microsoft in rare cases, feminine) grammatical gender is Translator displaying a lesser degree of bias, and DeepL used to refer to a generic or specific, but unknown, generally being more gender-neutral. person, whose actual gender cannot be guessed
Similarly to Costa-jussà [16], Sun et al. [18] conducted (e.g. “il vincitore riceverà un premio”, the.m.sg a comprehensive literature review, exploring various winner:m.sg will.receive a bonus, where the idenstrategies proposed in existing research to address gen- tity of the winner is unknown, and so is their der bias, including dataset preprocessing, algorithmic gender). modifications, and post-processing techniques. The paper emphasizes the significance of mitigating gender bias A second annotation layer concerns the proposal of in NLP systems and highlights the challenges associated gender-fair reformulations of the portions of texts identiwith bias detection and mitigation. ifed as unfair, 3 and the assignment of each reformulation
More recently, Stanczak and Augenstein [19] identify to a specific type. four key limitations in current research on gender bias in As for cases of incongruous gender, the only possible NLP. Firstly, social gender is often treated as a binary vari- type of gender-fair solution is providing a ‘congruous’ able, not paying attention to its fluidity and continuity. (It. “congrua”) alternative option, where the grammatiSecondly, studies usually give more importance to high- cal gender matches the gender of the referent (e.g., “la resource languages - primarily English - neglecting the diversity of languages spoken globally. Thirdly, despite 3At least one reformulation for each gender-unfair portion is required, but more than one reformulation is often provided.
visibility conservative innovative il vincitore o la vincitrice il vincitore o la vincitrice o l* vincitor* obscuration chi vincerà l* vincitor* Data Collection The data made available by the University of Brescia include a range of administrative materials such as the department’s strategic plan, reports from the departmental council and parity commission, as well as various forms. Most of them are already public on the website of the University, others, like the forms, have been asked to administrative organs. For this pioneering study, we collected specifically 13 documents.
illustrated in Table 1, where one example reformulation per type is provided for the phrase il vincitore (see the Appendix for other examples).
Lastly, ‘mixed’ (It. “ibride”) reformulations use diferent strategies for diferent elements in the gender-unfair portion of text, e.g. l* vincitore o vincitrice (the.inn.sg winner:m.sg or winner:f.sg), where an innovative obscuration strategy is used for the article and a conservative visibility strategy is used for the noun.
• ‘visibility’ (It. “visibilità”) strategies, that make
the possible reference to persons with diferent
genders explicit by means of the use of diferent
grammatical genders; vs. ‘obscuration’
(“oscuramento”) strategies, that try to avoid the use of
expressions that reveal the (assumed) gender of
referents; Data Preprocessing All the documents have been
• ‘conservative’ (It. “conservative”) strategies, that transformed into plain text to be processed
automationly use expressions that are part of the gram- cally. To deal with the special format of forms or the
matical system of the standard variety of Italian; layout in tables of special reports, we designed various
vs. ‘innovative’ (It. “innovative”) strategies, that regular expressions to clean and prepare the texts for the
introduce new means of expression into the sys- segmentation in sentences.
tem. To support the annotation phase, we signaled for each
sentence the possible presence of discrepancies:
displayed below each task were any words from the
enriched lexicon of professions’ names based on Sabatini
[
the annotation scheme, which was analysed during an initial meeting. Doubtful cases were discussed in regular 4. Corpus bi-weekly meetings. In addition, we kept a file of notes in which we reviewed and discussed uncertain cases as The scheme of annotation was applied by 5 expert an- they arose. The annotation scheme was subsequently notators of gender-fair language to a small corpus of ad- updated in response to the insights gathered both in the ministrative texts coming from the University of Brescia. file and the meetings. The annotation process has been Diferently from other textual genres, the administra- carried out on LabelStudio platform4, creating a specific tive texts, for their format and technical language, have interface that facilitates the two layers of annotation: the required a specific preprocessing process to let the an- identification of the gender-unfair expression, and the renotators focus especially on the spans of text that could formulation with one or more alternatives. The interface contain discrepancies. provided a section for comments in order to encourage
To this purpose, we employed the original lexicon of reflection on the annotation scheme and collect insights
professional names taken into account in Sabatini [
In the 422 sentences containing gender-unfair expres- expressions, but any element that needs to be changed in sions, the annotators detected on average 3 textual spans order to get a gender-fair text. For example, if an annoper sentence containing gender-unfair language (for a tator decided to propose "il corpo docenti" (the teaching total of 3,195 portions) and proposed from 1 to about 11 staf) instead of "i docenti" (the. m.pl teachers) in "i docenti alternatives. devono partecipare" (the teachers have to participate),
Moreover, looking at the frequencies of the types of they also have to select the verb "devono" for it to agree unfair expressions identified in the corpus, we can see in number with "il corpo docenti". Crucially, the verb from Table 2 that the most common case of gender-unfair only has to be selected if the proposed alternative to "i language in administrative documents is represented docenti" is singular and, thus, the verb needs to be sinby the use of overextended forms, and in particular of gular too. Hence, if another annotator doesn’t propose overextended masculine (e.g., “i ricercatori”, the.m.pl re- a singular alternative to "i docenti", they won’t need to searchers:m.pl; and “i docenti”, the.m.pl teachers, for select the verb. As a result, the two annotators would mixed-gender groups of, respectively, researchers and select diferent spans even when agreeing on what are teachers). the gender-unfair expressions within the text. For this reason, even comparing just the textual spans between #_portions type annotators would not be a good indicator of the annota2,709 overextended tors’ agreement. 452 generic However, in Table 3 below we provide an example of 34 incongruous a sentence with the reformulations proposed by the five 3,195 total annotators for each gender-unfair span of text, to give an idea of the kind of variation that can be found.
In recent years, gender-fair language has gained significant attention, leading to the proposal of new strategies in various languages to avoid using masculine or feminine endings. Motivated by these theories, we conducted a study to examine the usage of diferent solutions in practical situations. We developed guidelines for gender-fair annotation and reformulation of texts, which we applied to a corpus of 1,024 portions of university administrative documents in Italian.
The corpus was annotated by 5 experts, and in 422 cases the annotators identified at least one gender-unfair expression. The preliminary analysis of this corpus highlighted the need to adopt specific guidelines (as well as a list of words to pay particular attention to) to support administrative staf in writing gender-fair texts.
Applying our annotation and reformulation guidelines to real data has led to theoretical advancements: we discovered that ‘obscuration’ and ‘visibility’ strategies can coexist within the same reformulation, and we consequently updated the annotation scheme to include ‘mixed’
reliability has not been calculated for diferent reasons.
First, the scheme provides a variety of gender-fair options and the choice of a specific alternative depends on several factors, including individual preference: one annotator might agree on "l@ Professor@" (the Professor:INN.SG) being gender-fair, but choose a diferent innovative option, like "lu Professoru" (the Professor:INN.SG), instead.
Therefore, comparing the alternatives provided by each annotator is not a good measure of whether the annotators consider a specific option an appropriate gender-fair alternative to a certain gender-unfair expression. Relatedly, we do not plan to release an aggregated dataset, inclusive of a set of “gold-standard” preferred labels. Indeed, the aim of the present work is not to consolidate a ‘ground truth’ among rephrasing strategies but rather to explore as many solutions as possible while using genderfair language. This is tied to our focus on methodology:
Text: A partire dal 2013 il DiGi ha organizzato ogni anno International Summer schools, allo scopo di attrarre studenti stranieri e di ofrire agli studenti bresciani l’opportunità di entrare in contatto con studenti e docenti di altri Paesi. ann.
span
A B C D E bres– ‘di origine straniera’ or ‘di nazionalità estera’ a ‘studenti provenienti dalla provincia di Brescia’ ‘studenti di università estere’ ‘a coloro che studiano all’Università di Brescia’ ‘dall’estero’ – ‘dalla provincia di Brescia’ – – – – – – – ‘stranieri’ ‘agli studenti bresciani’ ‘studenti stranieri’ ‘agli studenti bresciani’ – – – – ‘studenti stranieri’ ‘studenti ciani’ ‘studenti’ ‘studenti/esse stranieri/e’ bres- ‘studenti/esse bresciani/e’ ‘studenti/esse’ ‘stranieri’ ‘agli studenti’ ‘bresciani’ ‘stranieri’ ‘agli studenti’ ‘bresciani’ ‘straniere/i’ ‘alle/agli denti’ ‘bresciane/i’ ‘straniere/i’ ‘alle/agli denti’ ‘bresciane/i’ stu– – – – – – – ‘straniere/i/3’ or ‘di nazionalità straniera’ – ‘straniere/i/u’ stu- ‘alle/agli/all3 studenti’ or ‘alle persone che studi- ‘all3 studenti’ or ‘alle/agli/allu studenti’ ano’ ‘all* studenti’ ‘bresciane/i/3’ or ‘bres- ‘del bresciano’, ‘di area – ciane/i/u’ bresciana’, or ‘di Brescia’
alternatives.
To summarize, the annotation scheme has proven valuable both practically and theoretically. It facilitated the The annotators have been paid in the context of the acidentification of gender-unfair expressions and the for- tions provided by the Gender Equality Plan of the Unimulation of alternatives. Moreover, it revealed the inad- versity of Brescia. The time of annotation has been monequacy of an exclusive distinction between obscuration itored to ensure that the actual time spent annotating and visibility, emphasizing the need to incorporate a new matched the agreed-upon paid hours. type of strategy (i.e., ‘mixed’ alternatives) into the classiifcation. Limitations
Although the created annotation scheme has been applied only to administrative texts so far, the guidelines are Our work presents some limitations. Firstly, the sample formulated in such a way that they can be easily applied of analyzed texts is small and related to a specific doto data pertaining to diferent domains. Indeed, we plan main. To test the robustness of the proposed guidelines, to extend the annotation to other data, like web pages of we planned to expand this corpus and its analysis. Seca University that describes its organization and its events. ondly, in this work we presented an annotation schema Finally, the resulting corpus, composed of 3,195 portions to recognize gender-unfair language and to reformulate of texts identified as gender-unfair and reformulated with it, specifically for Italian, limiting its adaptation to other at least one alternative, could be used in the context of languages. training models to recognize gender-unfair expressions and suggest their alternatives. Figure 1: Label Studio set up with a text containing an overextended and a generic gender-unfair text span. Figure 2: Label Studio set up with a text containing an incongruous gender-unfair text span.