Gender-fair translation: a case study beyond the binary

ABSTRACT

The increased visibility of non-binary individuals has fostered discussions around language and inclusivity. Due to differences in grammatical structures and language-specific strategies to represent non-binary genders, translation from notional gender languages into grammatical gender languages is extremely challenging. Hence, I propose a case study in gender-fair translation from English into German. Six language professionals translated three different texts. For each text, participants were instructed to use a different approach to gender-fair language, namely gender-neutral rewording, gender-inclusive characters, and neosystems. The focus of the present study is not on translation quality but rather on the cognitive processes and ease of integrating gender-fair language in the translation process. Findings from screen recordings show clear differences in translation times among participants and only partially among strategies. Results from retrospective interviews, non-participant observation, and target text annotations show increased keyboard activity and perceived difficulty in neosystems as well as lower success in their application.

KEYWORDS:

• Gender-fair language
• gender-fair translation
• queer translation
• non-binary
• linguistic inclusivity

1. Introduction

Non-binary individuals have seen an increased visibility in recent years, for example in English language TV series such as One Day at a Time (2017–2020) or Sex Education (2019– …). This has fostered discussions around inclusivity and language, especially which gender-fair language strategies are more appropriate to linguistically represent non-binary genders. In this paper, I use gender-fair language (GFL) as an umbrella term that subsumes two different approaches, gender-neutral and gender-inclusive (Sczesny et al., 2016). The first refers to strategies that conceal gender in language, e.g., the use of gender-neutral words such as person, passive constructions, and indefinite pronouns. The second describes the use of typographical characters, neomorphemes, and neopronouns to make non-binary or all genders visible. Due to grammatical differences among languages and a multitude of language-specific strategies, gender-fair translation is a challenging task for both humans and machines. Nevertheless, gender-fair language leads to greater visibility and equality for non-binary genders.

First, research on gender-fair translation usually concentrates on the analysis of audio-visual products or news articles in which non-binary individuals are portrayed, with a focus on instances of misgendering and/or the applied strategies (Attig, 2022; Lardelli & Gromann, 2023c; López, 2022). Human translators find it difficult to select and correctly apply gender-fair language strategies. Moreover, machine translation (MT) outputs are known for their gender bias (Savoldi et al., 2021). Therefore, I propose a case study to investigate the translation process of gender-fair language.

In this case study, I focused on the cognitive processes of translating texts featuring both non-binary as well as a multitude of genders. I also analysed the success of the integration of GFL in the translation process. Six language professionals translated three texts from English into German, which contain references to non-binary genders and mixed-gender groups. Whereas in English singular they has become predominant along with the use of gender-neutral nouns, in German there are many different approaches to GFL. Participants were instructed to apply one specific strategy per text corresponding to three main approaches, i.e., gender-neutral rewording, gender-inclusive characters, and neosystems. Screen recordings allow measuring the translation times. They also allow completing observational protocols from non-participant observation to reproduce the keyboard activity and hence the translation process. Retrospective interviews provide insights into the participants’ perceived difficulty and experience depending on the utilised GFL strategy. Finally, analyses of the produced translations reveal the success of the task.

Currently, there is no one-fits-all solution to both represent non-binary individuals as well as mix-gender groups and the strategy selection is extremely context-depended (Gromann et al., 2023). Even though (Piergentili et al., 2023) advocate for gender-neutral options, this may not always be the best solution. For instance, participants in the retrospective interviews commented that repetitions of proper nouns instead of third-person pronouns may lead the text to be perceived as redundant while neutral alternatives (e.g., ‘an acting person’) may be perceived as unnatural. In this study, professional translators expressed a clear preference for a combination of gender-neutral rewording with gender-inclusive characters over neosystems. With findings from this first case study, I contribute to guidelines for integrating gender-fair language into the translation workflow. I also show which specific approaches and gender-specific constituents are particularly challenging in source and target texts which could partially guide strategy selection.

2. Theoretical background

In order to provide a theoretical basis for gender-fair translation, this section introduces the relation between gender and language as well as gender-fair language strategies for English and German. The term gender has different meanings. It can refer to gender identity, i.e., the inner sense of one’s gender, and is opposed to sex assigned at birth (Barker & Iantaffi, 2019). In linguistics, it can refer to grammatical gender (Corbett, 1991). This is defined as “classes of nouns reflected the behaviour of associated words” (Hockett, 1958, p. 231) and means that word classes are inflected in agreement with the noun’s gender.

On the basis of their gender systems, languages can be categorised into grammatical gender, notional gender, and genderless languages (McConnell-Ginet, 2013; Stahlberg et al., 2007). In the first case, each noun has a gender, i.e., also animals, inanimate objects, and abstract concepts. Languages belonging to this category such as German are heavily marked for gender. In notional gender languages such as English, gender is mainly expressed in third-personal singular pronouns (he/she/it) and in nouns referring either to kinship (mother/father) or professions (waiter/waitress). In both categories, gender is also referential, i.e., common nouns in reference to people are inflected based on the extra-linguistic reality, i.e., the referent’s gender identity. Finally, genderless languages such as Turkish are mainly neutral.

These differences among language structures are also reflected in the multitude of gender-fair language-specific strategies. As already mentioned, English does not require extensive gender specification. Singular they has become predominant to refer to hypothetical people, people whose gender is unknown, and often to non-binary people (Baron, 2020). Furthermore, gender-neutral alternatives to nouns (e.g., chairperson instead of chairman or – woman) have been proposed. In German, four different approaches to gender-fair language can be identified: gender-neutral rewording, gender-inclusive characters, gender-neutral characters and forms, and gender-fair neosystems (see Lardelli & Gromann, 2023c for an in-depth overview). In the first case, sentences are structured to avoid gender marking, for example using gender-neutral words, indefinite pronouns, and/or passive constructions amongst others. Gender-inclusive characters such as gender star (*) are used to separate masculine forms of nouns and feminine endings as in ‘Leser*in’ (masculine*feminine reader). Similarly, gender-neutral characters or forms such as x as in ‘Lesx’ (reader, gender-neutral) are used to question the gender-binary and as a provocation. Finally, gender-fair neosystems introduce a fourth gender in the language through a set of new endings, pronouns, and articles. An example of such an approach is the ens-forms (Hornscheidt & Sammla, 2021) as in ‘Lesens’ (reader, neutral/non-binary).

Language usage can have a broader societal impact. For instance, Sczesny et al. (2016) found that the use of gender-fair language in contrast with masculine generics can reduce stereotyping and discrimination and contributes to the visibility of genders other than the masculine one. Furthermore, misgendering, i.e., the assignment of gendered forms that do not correspond to one’s gender identity, can cause emotional and psychological distress (McLemore, 2015). Translators can therefore become advocates for gender-fair language contributing to increasing gender equality not only for women but also for all genders beyond the binary.

3. Related work

Although feminist translation (Von Flotow, 2016) has a strong tradition, queer translation studies are relatively new (Baer & Kaindl, 2018), and language beyond the binary is rarely addressed in publications (Attig, 2020). Lardelli and Gromann (2023a) performed a first comprehensive literature review of gender-fair language in both translation studies and machine translation. Most of the research in the field analyses instances of misgendering and/or GFL strategies used in audio-visual products and media reports. This line of research is briefly introduced in the next paragraphs.

Attig (2022) and López (2022) analysed the dubbed and subtitled versions of the Netflix series One Day At A Time in Spanish and French. They found that instances of misgendering or identity erasure were quite common in the translations. Furthermore, translation strategies varied based on the version (subtitled vs. dubbed) and the language variety (European and Latin American Spanish). In three out of the four Spanish versions, the non-binary protagonist Elena was addressed with feminine forms, and/or a literal translation of singular they which is nonsensical for Spanish speakers. Only in the European Spanish dubbing, the non-binary pronoun ‘elle’ was used. In the French dubbing, the invented pronoun ‘zi/zu’, probably inspired by the English neopronoun ‘ze/zir’, was used along with the indefinite pronoun ‘on’ (one/we). In the subtitled version, ‘on’ was still used but this time along with the non-binary pronoun ‘ielle’.

Misiek (2020) observed a complete erasure of non-binary characters in the Polish translation of three different English-language TV series. Said characters were usually addressed with forms corresponding to their sex assigned at birth and their gender identity was not mentioned. Similarly, Šincek (2020) analysed Croatian movie translation and news reports on Sam Smith’s coming out as non-binary. The most frequent strategy was the use of the masculine plural noun, thus a literal translation of English singular they.

Lardelli and Gromann (2023c) created a corpus with articles on Demi Lovato’s coming out as non-binary in Italian and German. While gender-fair language was actively used in a few cases (e.g., non-binary German pronoun ‘xier’ was found once in the corpus), instances of misgendering as well as literal translations of English singular they were extremely frequent.

4. Method

The present method is inspired by Translation Process Research (TPR) Jakobsen (2017) and Albl-Mikasa et al. (2017). It combines screen recordings, non-participant observation, retrospective interviews, and translation annotation. Six professional translators with at least three years of work experience were recruited.¹ Prior to the study, they completed a questionnaire on their profile as well as gender-fair language use and impressions. They received instructions on the tasks and a handout on gender-fair German that they could use to prepare for and during the study.

As shown in Table 1, participants received three texts of approximately 150 words each on three different English language TV series, i.e., Sex Education, Grey’s Anatomy, and Sort Of.² Texts were retrieved online and adapted for the study. They discussed non-binary actors joining said TV series. To ensure comparability of translation times per text, readability scores were computed using the Flesch-Kincaid readability test (Kincaid et al., 1975). This test considers the number and length of words, but it ignores semantics. The texts contained references to non-binary individuals and mixed-gender groups. German was selected as the target language because it requires extensive gender marking, e.g., in pronouns, nouns, and articles, compared with English. Besides, a multitude of gender-fair language strategies have been proposed and they are quite different in their use.

Table 1. Details on the translation materials.

Text no.	TV series	Instructed gender-fair approach	Word count	Gendered phrases
1	Sex Education	Gender-Neutral Rewording	152	9
2	Grey’s Anatomy	Gender-Inclusive Characters	151	12
3	Sort Of	Neosystems	163	10

Translators received a text file containing the assignments. They could decide whether to use the delivered file for the translation or import it into their CAT tools. The use of machine translation was not permitted. Each text corresponded to a different assignment for which a different approach to gender-fair language, namely, gender-neutral rewording, gender-inclusive characters, and neosystems, was to be used. Participants could freely select specific strategies described in the provided handout for each approach. For example, they could decide whether to use gender star (*) or underscore (_) in the case of the second assignment.

The study was conducted online because it aimed for a most authentic and unintrusive experimental setting. Translators could work in their familiar environment and were instructed to work under usual conditions. However, they were required to work on one screen only which was recorded during a videoconference open in the background. This ensured that the whole translation process was recorded. After the study, participants were interviewed about the strategies used, their impressions, the challenges faced as well as possible use of machine translation for similar assignments. The interviews were conducted in German, transcribed according to semantic transcription rules (Dresing & Pehl, 2018), and analysed with quantitative content analysis (Kuckartz, 2014) using qualitative analysis software MAXQDA.

I drew on Krings’ (2001) division into temporal, technical, and cognitive effort to analyse cognitive processes of translating gender-fair language. Even though this categorisation is often utilised in post-editing studies, it can be useful to investigate the gender-fair translation process from different angles. Screen recordings were used to measure translation times. Together with observational notes from non-participant observation, they were also used to reproduce and analyse the keyboard activity. Based on Muñoz and Cardona Guerra (2019, p. 536), broken units, changes, pauses (3 + seconds), and searches were used as indicators of potential translation problems leading to an increased keyboard activity. Retrospective interviews were used to gain insights into the participants' perceived difficulty and personal experiences with the tasks. Finally, gendered phrases in the translation were annotated for the GFL strategies selected and the success of their use. For each text, respectively nine, twelve, and ten gendered phrases were identified (s. Table 1). They were composed of different word classes, such as nouns, pronouns, adjectives, and articles.

5. Results

First, the participants’ profile is presented. Then, the temporal, technical, and cognitive effort for each text is illustrated. Finally, an overview of the strategies used as well as the success of their application is provided.

5.1. Participants

As shown in Table 2, participants were translation professionals with work experience spanning from 3–5 to 11–15 years and they were all aged around 30–40 years old. Four identified as women, one as a man, and one as non-binary. All participants had already some knowledge of gender-fair language and all use it in their daily work. Most of them use gender-inclusive characters, whereas one person indicated selecting the strategies depending on the client and/or assignment. Finally, one person indicated using all strategies.

Table 2. Participants’ profile.

Participant	Age	Gender	Work experience	GFL experience	GFL use
P1	39–45	Woman	11–15	Little	Depends on client/assignment
P2	25–31	Woman	3–5	Yes	Gender-inclusive characters
P3	25–31	Man	3–5	Little	Gender-inclusive characters
P4	32–38	Non-binary	6–10	Yes	All
P5	32–38	Woman	6–10	Yes	Gender-inclusive characters
P6	32–38	Woman	6–10	Little	Gender-inclusive characters

Reasons for GFL use are mostly related to feminist beliefs and the promotion of gender equality. As regards challenging aspects of GFL use, they mentioned: (i) impact on the text’s readability, (ii) difficulty in finding neutral alternatives to gender-specific nouns and consistent use, and (iii) strategy selection based on the specific context, amongst others. Participants were also required to rate GFL difficulty on a Likert scale from 1 to 5 where 1 stands for very difficult and 5 for very easy. Most participants indicated neutral to difficult, as shown in Figure 1.

Figure 1. GFL difficulty.

5.2. Ttranslation times

Translation times per each assignment were similar. As shown in Figure 2, there were considerable differences in translation speed among participants. Hence, the median is here used as a better value of the data’s central distribution. Participants needed 00:28:11 min (SD = 00:04:49) to complete the first assignment, 00:27:48 min (SD = 00:06:23) for the second, and 00:33:21 min (SD = 00:08:39) for the third. Thus, there was no difference in translation times between gender-neutral rewording and gender-inclusive characters. Neosystems required the highest temporal effort. However, a repeated measures ANOVA did not show a statistically significant difference among assignments (F[2,10] = 0.674, p-value = 0.531). When looking at the translation times for each participant, it can be noted that most translators were similarly fast in each assignment except for P2, P5, and P6. The first two needed much more time to complete the text with neosystems compared with the other assignments whereas P6 was slower in gender-neutral rewording.

Figure 2. Translation times in minutes.

In order to better compare translation times among assignments, measurements were also normalised. Translation times for each task were divided by the number of words in the source texts as shown in Table 3. Marginal differences among assignments are found whereas the standard deviation was quite high. This confirms that there were considerable differences among participants in terms of translation speed, especially in the third assignment. P4, for instance, was faster in this assignment than in the first two whereas temporal effort was very high for P2 and P5 when compared with gender-neutral rewording and gender-inclusive characters. Besides, participants (P1, P3, P4, P5) were usually faster in the first assignment than in the second. In general, P4 and P5 were the fastest participants and needed between 18 and 25 min for each text. P3 was generally the slowest translator with about 34 min per text, except for P2 in the third assignment who needed about 40 min.

Table 3. Median value of the translation times, standard deviation (second per word), and relative standard deviation by text.

Assignment	Times (s/w)	Relative SD
1	10.9 ± 1.9	17%
2	11.2 ± 2.6	23%
3	13.2 ± 2.8	24%

5.3. Keyboard activity

The analysis of the keyboard activity shows clear differences between the first two assignments and neosystems. Table 4 shows the total occurrence and the median of each indicator of technical effort (broken units, changes, searches, and pauses) as well as their sum. Gender-neutral rewording (n = 159, Mdn = 27, SD = 5.3) involved a higher keyboard activity than gender-inclusive characters (n = 133, Mdn = 20.5, SD = 10). Neosystems (n = 234, Mdn = 41, SD = 18.4) were the most difficult assignment. Standard deviation shows that there were great differences among participants, most in the case of neosystems. A repeated measures ANOVA with Greenhouse-Geisser correction, however, did not show a statistically significant difference in the keyboard activity among assignments (F[1.62,8.10] = 2.49, p-value  = 0.148).

Table 4. Overview of the technical effort for each assignment.

	1st Assignment		2nd Assignment		3rd Assignment
	Total	Median	Total	Median	Total	Median
Indicators	159	27	133	20.5	234	41
Broken Units	14	2	6	0.5	12	1.5
Changes	56	6.5	56	8	87	18
Searches	36	4.5	28	4	78	25.5
Pauses	53	10	43	5	57	2.5

The use of neosystems involved a higher number of changes in the translation (Mdn = 18) in comparison with gender-neutral rewording and gender-inclusive characters (Mdn = 6.5 and 8 respectively). This difference was greater in the case of searches. These were far more frequent in the third assignment (Mdn = 25.5) than in the first two (respectively Mdn = 4.5 and Mdn = 4). Little differences among GFL approaches were found in terms of broken which were found quite sporadically. Pauses were more frequent in the first assignment (Mdn = 10) than in the second (Mdn = 5) and in the third (Mdn = 2.5).

As in the case of translation times, differences were found among participants. Table 5 shows the total occurrences of indicators of possible translation problems per participant for each assignment. In the first assignment, participants face a similar number of difficulties, i.e., about 30. P5 was the exception with only 18. In the second assignment, the keyboard activity widely differed among participants, spanning from nine (P3) to 37 (P5) instances of technical effort. This time, P5 faced a higher number of difficulties than the other participants. Differences among participants were more considerable in the third assignment. P4 stands out with only 15 instances of translation problems whereas P3 faced the highest number of difficulties (n = 61).

Table 5. Technical effort per participant for each assignment.

Participant	1st Assignment	2nd Assignment	3rd Assignment
1	29	30	30
2	24	21	52
3	30	9	61
4	33	20	15
5	18	37	52
6	25	16	24

To understand which gender-specific constituents caused the greatest translation problems, a more fine-grained analysis of the keyboard activity was performed by dividing each source text into seven segments. Segments contained a different number of gendered instances. Due to space constraints, we present here a selection of translation problems for each text.

In the first text, the following segment caused great difficulties: ‘Leading the way is a non-binary actor and musician Dua Saleh as Cal, a new student at Moordale who clashes directly with Headmistress Hope’. P1, for example, translated student as ‘Neuzugang’ (EN: new entry). Then they made a nine-second pause and looked online for this term and ‘Neuankömmling’ (EN: newcomer). Then, they made another pause, this time of ten seconds. They changed the translation from ‘Neuzugang’ into ‘Neuankömmling’. They deleted this last translation, made a six-second pause, and then opted for ‘neuen Mitglied’ (EN: new member). While translating the same segment, P2 looked four times in the provided handout for inspiration. Also, before finding a solution for ‘actor and musician’, they made two long pauses, respectively of 38 and ten seconds. Then they started writing ‘Schaus’ (EN: act), made another pause of nine seconds, wrote ‘Schauspiel’ (EN: acting), deleted it, and wrote ‘Schauspielperson and Musik’ (EN: acting person and music). They did not complete the translation but made a 21-second pause. Finally, they opted for ‘Musikmensch’ (EN: music person).

Most of the problems in the second text were contained in the first segment: ‘Grey’s Anatomy has its first non-binary doctor after E. R. Fightmaster was promoted from a recurring guest star to a recurring cast member’. P5, for instance, started translating such a phrase using masculine forms – ‘Grey’s Anatomy hat seinen ersten nicht-binären Arzt’. Then they tried to apply gender star to the translation, thus writing ‘seine*n erste*n’ (EN: its first) and making a 19-second pause. They reverted the translation to the original solution with masculine forms and highlighted it. Afterwards, they tried to add an alternative translation and started writing ‘/sein’ (EN: /its), but deleted it. They re-translated the phrase by using feminine forms, hence ‘/seine erste nicht-binäre Ärztin’ and highlighted this alternative as well. In brackets, they added another tentative translation: ‘seine*n erste*n nicht-binären’ (EN: its first non-binary). Then they made a 14-second pause, decided to use star also for the adjective ‘non-binary’, and opted for a synonym instead of doctor: ‘(seine*n erste*n nicht-binäre*n Mediziner*in)’ (EN: its first non-binary physician). While revising the whole translation, P5 also searched online how to use gender star with the word ‘Arzt’ (EN: masculine doctor) because the feminine form requires umlaut in the word stem (Ärztin, EN: feminine doctor).

In the third text, the most problematic segment was the third, namely ‘This is a comedy about a gender-fluid nanny navigating their complicated existence by co-creators Bilal Baig and Fab Filippo’. Specifically, German equivalents to nanny are all of feminine gender, complicating the translation process. P6, for example, looked in an online bilingual dictionary for possible solutions. They then searched for ‘Kinderfrau’ (a synonym for nanny) in a German monolingual dictionary. Finally, they translated the term as ‘Kinderpflegens’ (EN: baby-minder, gender-neutral) and decided to delete it and use ‘Kinderbetreuens’ instead, which is another synonym.

5.4. Perceived difficulty and personal experience

During the retrospective interviews, participants discussed each text and strategy commenting on their solutions, difficulties, and personal preferences. They also elaborated on their general experience as translators in the context of the study and on the opportunity to utilise machine translation for similar tasks.

In selecting the easiest assignment, participants were equally divided between gender-neutral rewording and gender-inclusive characters. Nevertheless, such GFL approaches were also selected as the most difficult assignment by one participant respectively. Neosystems were never regarded as the easiest assignment, but only two translators explicitly referred to it as the most difficult.

Attitudes towards gender-neutral rewording were contrasting. Even though half of the participants found it the easiest assignment, the other half indicated that great effort is required when utilising it. While comparing the assignments, one participant commented that ‘they did not require as much brain capacity as the first text’. Half of the translators indicated the need to be creative, which was both seen as positive and negative. One participant noted that ‘this assignment was the easiest because I could use my own language repertoire’, whereas another found it ‘time-consuming and complicated’. Furthermore, there seems to be quite an agreement (n = 3) that the target text reads differently, for example, because it ‘becomes very impersonal’, e.g., due to the use of passive constructions to avoid gender. They also commented that ‘one is constrained in the wording’, and sometimes repetitions of nouns are needed to avoid gendered pronouns. Finally, participants indicated that it is difficult to find neutral alternatives to commonly gendered terms, as in the case of ‘students’ in the English source text. The most difficult gender-specific words to translate were ‘actor and musician’ (n = 4) as well as the ‘student’ (n = 3).

Even though the use of gender-inclusive characters was regarded as generally easy, participants commented on many negative aspects of this approach. The vast majority (n = 5) found it difficult to ensure consistency in how the characters are used, e.g., ‘should it be sie:er (EN: she*he) or er:sie (EN: he*she)? I decided to use the masculine form first […] because then I had the word ‘Zuseher*innen’ (EN: viewers) […] and hence it would have been consistent’. According to four translators, the sole use of characters, without also opting for rewording/and or neutral words affects readability because the text ‘becomes a hail of characters’. There was quite a consensus (n = 4) that this approach should be utilised along with gender-neutral rewording and that using a strategy only constrains the translators (n = 4). Another interesting aspect is that five participants stated they need to think of both masculine and feminine forms of words to combine them with characters. Hence, this approach challenges the gender binary only partially. Finally, two translators specifically identified the translation of ‘doctor’ as difficult because the feminine form differs from the masculine in the word stem as well (Arzt vs. Ärztin).

The use of neosystems presented numerous difficulties and was described as ‘requiring so much brain capacity […] that I had less capacity to find other nice wordings’. Participants (n=4) noted that such an approach ‘requires further training […] but it can be used’ even though ‘it would be weird’. In this case, translators (n=2) were not sure about whether they were correctly applying the selected system and they also needed to use the provided handout for the entire translation process. The greatest difficulties concerned the use of adjectives (n=4) with one participant not being sure which adjectives should be gendered anymore: ‘genderfluid – I would probably decline the adjective as well because it refers to the genderfluid nanny. However, I did not decline ‘complicated existence’ (another adjective found in the text)’. Finally, the term ‘nanny’ was challenging for quite each participant (n=5), who for example decided to use ‘Kinderbetreuens (EN: caregiver) first, but then I thought that nanny is quite common in language use’.

When asked to comment on their use of machine translation, participants revealed quite a negative attitude toward this technology. Half of them indicated not using it at all, whereas one uses it sometimes as a source of inspiration to find more translation solutions. Most of the interviewed translators (n = 4) would not use MT for similar texts, indicating extensive post-editing required to adapt the text style (n = 5) and the lack of gender-fair language (n = 5).

Translators (n = 4) finally expressed the desire to receive clearer instructions from their clients as to which gender-fair language approach is preferred, also noting (n = 1) that general awareness on the topic is still needed. Finally, they (n = 3) would also need more literature or gender-fair language dictionaries to facilitate their work.

5.5 Strategy selection

For each assignment, participants could freely select strategies from the provided handout or other known strategies. Gender-neutral rewording is a creative approach that can be realised differently, e.g., gender-neutral words, passive constructions, name repetitions, omissions of pronouns, and sentence rewording. Gender-inclusive characters are relatively straightforward as an approach because typographical characters are used to separate masculine and feminine forms of gendered words. Characters can include star (*), underscore (_), and colon (:). Finally, a multitude of neosystems has been proposed and they can range from very simple ones, such as the ens-Forms which are used as pronouns and gendered endings, to more complete and complex ones, such as the Sylvain System which introduces new pronouns, gendered endings, and declension system.

While the use of specific strategies was found inconsistent among participants in their application, the number of mistakes differed based on the assignment. The term mistake as used for the present analysis includes misgendering, grammatical mistakes in the use of gender-fair language, translation of singular they in the plural, and use of wrong terms. While the number of mistakes was low in the first and second assignments (respectively 17% and 14% of the analysed gendered phrases), this value increased in the third (36%).

In the first assignment, participants had to be very creative to avoid gendered constructions, hence they combined different strategies. The translated source text contained nine gendered phrases of interest for the present analysis. For the 54 analysed phrases, 55 annotations for strategies were performed. Table 6 shows the strategies applied and their frequency as well as the type and number of mistakes found in the translated phrases. Gender-neutral words and rewording were the most common strategies (respectively 22% and 20% of annotations). Some examples include ‘nicht-binäre Schauspiel – und Musiktalent’ (EN: non-binary talent in cinema and music) or ‘nicht-binäre Person, die für ihre Musik bekannt ist’ (EN: non-binary person, who is famous for their music) as translations for the phrase ‘non-binary actor and musician’. Pronoun omission (16%), participles (13%), and name repetitions (7%) were also quite frequent, e.g., ‘Cal möchte eine lockere Uniform tragen’ (EN: Cal would like to dress a loose uniform) instead of ‘Cal likes to keep their uniform loose’. Collective nouns and relative pronouns were found just once (2%). A gender-inclusive character was also used even though this was not permitted. In 16% of cases, a gender-fair language strategy was not needed. Since the pronoun they was sometimes used in the source text in reference to mixed-gender groups, the German third-person plural pronoun ‘sie’ could be normally used. Only in three of the 54 analysed phrases (6%), there were instances of misgendering while in six (11%) the English noun ‘student/s’ was wrongly translated as ‘Studierende’ which in German refers to university and not high-school students.

Table 6. Strategies used and success of their application in the first assignment.

Strategy	N	%
Gender-neutral words	12	22%
Rewording	11	20%
Pronoun omission	9	16%
Participles	7	13%
Name repetition	4	7%
Collective noun	1	2%
Relative pronoun	1	2%
Gender-inclusive character	1	2%
No strategy	9	16%
Total	55	100%
Mistakes	N	%
Misgendering	3	6%
False Term	6	11%
Total	9/54	17%

The second text contained 12 gendered phrases of interest for the current analysis. As shown in Table 7, for the 72 translations, a total of 81 annotations were performed, i.e., in nine cases gender-inclusive characters were combined with a rewording strategy from the first assignment. Even though gender-inclusive characters were to be used, in a great number of cases translators opted for rewording sentences to avoid their use (35%) and, as in the first assignment, gender-fair language was sometimes not needed (17%). In the 72 analysed phrases, ten mistakes were found. Four concerned a misunderstanding of the singular they which was translated in the plural. Another four concerned the use of both masculine and feminine forms separated by a character for the translation of doctor (e.g., Arzt:Ärztin). An instance of misgendering was found and in one case a translator made a mistake in the adjective declension when using gender star.

Table 7. Strategies used and success of their application in the second assignment.

Strategy	N	%
Gender-inclusive characters	39	48%
Rewording	28	35%
No strategy	14	17%
Total	81	100%
Mistakes	N	%
Singular they	4	6%
Use of characters	1	2%
Misgendering	1	2%
Double forms	4	6%
Total	10/72	14%

Gender star was used by half of the participants although its application differed:

• One participant always used masculine forms first and then feminine forms separated by the star, e.g., ‘der*die Schauspieler*in’ (EN: masculine*feminine article masculine*feminine noun);

• Another one used feminine forms first in articles and pronouns, e.g., ‘die*der Schauspieler*in’ (EN: feminine*masculine article masculine*feminine noun);

• The last one used feminine forms first in articles and pronouns but combined in a new form, e.g., ‘die*r Schauspieler*in’.

Two other participants used underscore and masculine forms always first. However, one used the slash for articles and pronouns, as in von ‘dem/der brillanten Neurowissenschaflter_in’ (EN: of masculine/feminine article brilliant masculine_feminine noun). Finally, the last translators used colon.

The third text contained ten gendered phrases and a total of 63 annotations were performed. This means that translators combined neosystems with a rewording strategy three times. As shown in Table 8, neosystems were generally used (72%), but sometimes translators opted for rewording the phrases (22%). In 6% of cases, they did not need or opted to use gender-fair language. As the number of mistakes (37%) shows, the third assignment was the most difficult, with misgendering occurring four times and mistakes in the application of neosystems 18.

Table 8. Strategies used and success of their application in the second assignment.

Strategy	N	%
Neosystem	45	72%
Rewording	14	22%
No strategy	4	6%
Total	63	100%
Mistakes	N	%
Misgendering	4	7%
Use of neosystems	18	30%
Total	22/60	37%

Most of the participants (n = 4) selected the ens-forms, e.g., ‘einens südasiatisch, queer, muslimisch Schauspielens’ (EN: a South Asian, queer, Muslim actor). They (n = 3) indicated that this choice was due to the simplicity and intuitiveness of the strategy. Another participant utilised the Sylvain system (e.g., ‘einin genderfluidin Nanny’, EN: a genderfluid nanny) because of its completeness. Finally, the last participant opted for a system that they use in their daily life and consists of the use of ‘I’ as a gender-neutral ending and the non-binary pronoun ‘dey’ (e.g., ‘ein südasiatisches, queeres, muslimisches Schauspieli’, EN: a South Asian, queer, Muslim actor).

6. Conclusion & discussion

In this first gender-fair translation study, six language professionals translated texts containing references to non-binary individuals from English into German. A substantial variation in selecting gender-fair language strategies could be observed. When instructed to use gender-neutral rewording, participants opted for gender-neutral words, omitted pronouns or repeated the character/actor’s name. They also had to be very creative and reword whole sentences to avoid gender markers. When required to use gender-inclusive characters, half of the participants selected gender star (*). Nevertheless, its application was inconsistent, including a different order of feminine and masculine forms as well as the use of other, unconventional, characters such as slash (/). When instructed to use neosystems, most of the participants selected the ens-forms due to their simplicity. The variation in the use of strategies will most probably always occurs since there are many ways to reword a phrase and, as seen in the second assignment, there are no established conventions regarding how to use gender-inclusive characters.

A higher occurrence of mistakes in the application of neosystems was found in comparison with the first two assignments which clearly indicates an increased cognitive effort. This is also confirmed by the analysis of the keyboard activity, where searches and changes occurred more frequently. Interestingly, translators made more pauses while utilising gender-neutral rewording, probably because they had to think of neutral alternative to commonly gendered terms. In terms of translation times, however, the use of different strategies did not impact the participants’ productivity. Differences depended more on the person than the specific gender-fair language strategy.

In terms of perception, participants found gender-neutral rewording and gender-inclusive characters as easy approaches to gender-fair language. There were ambivalent opinions regarding gender-neutral rewording as it requires creativity, and this is sometimes perceived as challenging. The main pitfalls of gender-inclusive characters include the difficulty to ensure consistency in their application, also confirmed by the variation in use among participants, and the impact on readability. Furthermore, this strategy may not be suitable to question and/or dismantle the gender binary as participants indicated to thinking of feminine and masculine forms to apply it. Translators expressed the desire to combine these two approaches, which they also partially did during the translation experiments. This confirms how translation is a creative process that requires larger degrees of adaptation to the target language and cultural conventions. Neosystems were generally found difficult because they are largely unknown and feel like a foreign language, hence requiring further training to be used. This is confirmed by the number of mistakes found in the translation and shows how higher keyboard activity and increased perceived difficulty do not necessarily correspond to longer translation times.

A study with a larger sample of participants could shed light on this phenomenon and reduce the influence of personal differences in translation speed. Furthermore, since participants had a rather positive attitude toward gender-fair language and use it in their everyday work life, the replication of the study with a large and more varied population might lead to different results. This limitation has to do with the difficulty to find enough language professionals who were not recruited in a single, limited area and explains the online setting for the experiment (together with striving for experimental realism). Thus, the results obtained with the proposed mix of methods were interesting, but eye-tracking and key logging could be potential alternatives to provide a less subjective and more detailed evaluation of the cognitive effort.

Finally, participants were also asked to comment on the possible use of machine translation to investigate whether gender-fair translation might be a feasible method to provide datasets for a potential debiasing of MT. A rather negative attitude towards this technology seems to be still prevailing with participants lamenting the need for extensive post-editing (Lardelli & Gromann, 2023b) to produce a stylistic appropriate translation and to correct gendered references. Thus, further studies involving post-editing and comparing it to translation from scratch could provide insights into the effectiveness of MT within the context of gender-fair language use in the translation process. The language-specificity of GFL has also implications for machine translation. While there are first approaches to MT debiasing (e.g., Vanmassenhove et al., 2018), these mostly focus on binary genders and on the technical side of the issue of gender bias in MT. Participatory research (e.g., Gromann et al., 2023) may lead to community-informed debiasing approaches that consider the perspectives of different stakeholders such as non-binary people and language professionals.

In a nutshell, the results of this first experiment suggest a substantial variation in the gender-fair language strategy selection and use with a generally high cognitive effort. While utilising neosystems do not negatively impact translation times, results in terms of keyboard activity are more difficult to interpret. The number of searches and changes considerably increases but when all indicators of possible translation problems are considered, there are no statistically significant differences among assignments. However, participants clearly feel overwhelmed by the use of neosystems. Findings also confirm that there is no one-fits-all solution to the issue of representing non-binary people: a combination of different approaches is and may be preferred, and the strategy selection is highly dependent on numerous factors (e.g., target public, client preferences, and people mentioned in the texts).

Acknowledgements

I would like to warmly thank all the participants in the study. I would also like to thank my supervisors, Stefan Baumgarten and Dagmar Gromann, for their support while preparing and conducting my research. Finally, I want to thank the anonymous reviewers whose comments helped improve the final manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the UNIVERSITAS Austria.

Notes on contributors

Manuel Lardelli

Manuel Lardelli is currently doctoral candidate at the University of Graz’s Department for Translation Studies. His research focuses on the translation and post-editing of non-binary gender-fair language as well as on gender bias in machine translation and its socio-technological impact. He was also a research associate at the University of Vienna for the GenderFairMT project. He co-authored six peer-reviewed publications and was program committee member of the 1st International Workshop on Gender-Inclusive Translation Technologies (GITT 2023).

Notes

1 In compliance with regulations, ethic approval was sought and received for the present study – Ethikantrag Nr. 87 - 2021/22 in the manuscript with author details.

2 A link (https://doi.org/10.5281/zenodo.8413019) to the instructions and the study materials will be provided after the anonymity period.