This contribution is based on a presentation given at The Digital Orientalist’s Virtual Conference 2025 (AI and the Digital Humanities) by Moeka Kiyohara (University of East Anglia). The recording of the presentation can be found here.
The Japanese language features speech styles associated with specific gender groups, such as otoko kotoba 男ことば (masculine speech) and onna kotoba 女ことば (feminine speech). While real-life Japanese speakers do not commonly employ these gendered speech styles, they frequently appear in fictional works and translated texts (Nakamura, 2005; Furukawa, 2024). Nakamura (2013) identifies onna kotoba as an ideological construct shaped through discourse, emphasising the role of translation in sustaining gender ideologies in Japan.
Women’s speech in English is characterised by the tendency of language use, which is relatively unlikely to appear in the AI translation process (Lakoff, 1973). Meanwhile, the representation of gender differences in Japanese extends to stylistic features, creating a situation where AI models determine whether gendered speech should be incorporated in the translated text. Using case studies of English-to-Japanese translation, this study aims to introduce the structure of Japanese gendered language and how such specific gendered concepts embedded in language can pose challenges for AI translation.
Gender Differences in Japanese
Concepts of gender differ from one cultural group to another and are reflected in linguistic norms in different ways. Consequently, studies of language and gender should not treat gender as an isolated linguistic category, but should instead consider the social and cultural environment in which the language is spoken (Tanaka, 2014). The Japanese language features speech styles for the use by specific sociocultural groups to which the speaker belongs, namely yakuwarigo 役割語 (“role language”) (Kinsui, 2003).
Yakuwarigo encompasses various linguistic features such as personal pronouns, verb endings, and sentence-final particles that can indicate the speaker’s gender, age, social status, and regional background. As Kinsui describes, yakuwarigo is primarily used in fictional contexts rather than in everyday communication (Kinsui, 2003; Nakamura, 2013; Okamoto, 2010). Rather than reflecting natural speech, yakuwarigo functions as a linguistic marker that draws on the hearer’s knowledge of stereotypes.
While yakuwarigo offers writers an efficient way to signal a character’s identity without providing details, it simultaneously presents challenges for both writers and translators. The feminine speech is associated with linguistic forms conveying deference and softness, whereas the masculine speech employs forms expressing assertion and confidence. The feminine style, by its structure, compels speakers to employ more polite or indirect expressions, thereby imposing pragmatic constraints on women’s speech (Usami, 2003).
Gendered Yakuwarigo in Translation
Notably, Nakamura (2013) identified onnna kotoba as an ideological construct historically shaped through discourse, highlighting the contribution of translated speech in sustaining traditional gender ideologies in Japan. Past research has shown that real-life female Japanese speakers use gendered language less frequently than fictional characters do, and often alternate between masculine and feminine styles depending on their communicative goals (Okamoto and Sato, 1992; Furukawa, 2024). Meanwhile, in translated fiction, female characters are often assigned overtly feminine language that would be unnatural in real-life speech. Nakamura attributes this discrepancy to the intertextuality of translation, which enables the separation of the “body” and “voice” of a character. When reading translated texts, audiences perceive both the characters and the translator’s voices, making the projection of Japanese gender stereotypes onto non-Japanese characters appear naturalised. Thus, translated texts, through the bodies of non-Japanese women, can sustain and reinforce Japanese gender ideologies. This phenomenon highlights that translation is not a neutral act of linguistic substitution but a site where ideological constructs are negotiated and reproduced.
Challenges with AI Translation
Research on gender differences in English has shown that distinctions tend to appear in pragmatic features, such as hedges and tag questions, rather than as stylistic features like the case of yakuwarigo (Lakoff, 1973). Such differences highlight an ethical challenge for AI translation systems: in the case of English-to-Japanese translation, the AI model is responsible for the decision on whether gendered speech should be incorporated.
Because neural machine translation (NMT) systems rely on statistical probability derived from large datasets, they risk reinforcing dominant linguistic patterns that often contain gender bias. Furthermore, NMT models are typically trained on written language, which tends to display more stereotypical gendered features than spoken language in the case of Japanese. Recent studies also show the concern that training data for NMT is based on a fixed social relationship between interlocutors, overlooking the interpersonal aspects of communication (Hanari, 2019).
The issue can be illustrated through the translation of literary works, such as Hitomi Kanehara’s Autofiction (2007). In the original Japanese, the protagonist uses masculine features such as the pronoun omae お前 (you) and the sentence-ending daro だろ (I guess/right?) to express frustration. The English translation by Karashima maintains this confrontational tone, whereas the AI-generated translation via DeepL replaces these elements with soft, feminine endings such as –wa –わ and –ne –ね. The AI system’s lack of recognition of the strategic use of gendered forms underscores its limitations in interpreting communicative intent and social nuance.
| English Translation | DeepL Translation |
| Right away, I know what she’s up to. You have your eyes on my husband, don’t you, you bitch? I bet she spilled that champagne on purpose. So she could wipe his knee with the towel. | すぐに、彼女が何をしようとしているのかわかった。私の夫を狙っているんでしょ、このビッチ。きっとわざとシャンパンをこぼしたんだわ。彼の膝をタオルで拭くためにね。 |
Potential Humanities Approaches for AI Translation Research
Humanities-based approaches may provide valuable frameworks for addressing such limitations. One potential approach is the function-based model, which shifts analytical attention from what is translated to why it is translated. Using Text Typology Theory (Reiss, 1977/89), translations can be analysed based on the text type of the source text, such as informative, expressive, and operative texts. This approach sees the source text as a whole, paying attention to the ‘communicative achievement’ of the text rather than its linguistic content. Similarly, the Skopos theory (Vermeer, 1989/2004) considers the overall “purpose” (Skopos) of the translation. Applying these theories in the process of translating gendered language may involve posing questions, such as whether the translation’s purpose is to convey the speaker’s gender identity or to exaggerate the expression for entertainment purposes. However, applying function-based approaches to AI translation can pose challenges as the purpose of the translation is not always clear.
Hanari (2019) discusses the application of Politeness Theory (Brown and Levinson, 1987) to AI translation. Politeness is a complementary concept for the discussion of gendered yakuwarigo, as it is often discussed in relation to language and gender (Mills, 2003). Hanari argues that in order to apply Politeness theory to machine translation, it should be considered at the level of discourse, taking into account concepts like politic behavior (Watts, 2003) and unmarked politeness (Usami, 1999).
Conclusion
This study aimed to highlight the complex intersection between AI translation, gender, and culture, using Japanese gender yakuwarigo as a case study. As past research has shown, AI translation technologies reflect ideological stereotypes about gender due to its operation on statistical correlations. As a result, they may overlook social factors of communication, such as speakers’ social relationships and communicative goals. As we have discussed, the challenge is not merely to improve translation technologies, but to re-imagine how we conceptualise translation from cultural perspectives. By exploring potential humanities-based approaches, I hope to have highlighted the significance of considering language- and culture-specific constructions of gender into the growing body of AI and translation research.
References
Brown, Penelope and Levinson, Stephen. Politeness: Some Universals in Language Usage. Cambridge: Cambridge University Press, 1987.
Furukawa, Hiroko. Gender. Tokyo: Chikuma Shobō, 2024.
Hanari, Takushi. “Kikai hon’yaku to poraitonesu: Kikai hon’yaku ni hanei saseru beki poraitonesu to sono shuhō ni kansuru ichikōsatsu” 機械翻訳とポライトネス：機械翻訳に反映させるべきポライトネスとその手法に関する一考察 [Machine Translation and Politeness: A Study on Politeness Appropriate for Machine Translation and How to Realize it]. In Machine Translation and Future Society: Will We See a World without Language Barriers in the Future?, edited by Yasushi Takita and Yū Nishijima. Tokyo: Shakai Hyōronsha, 2019.
Happa no Kōfu. “AI hon’yaku o kenshō shitemita fittujerarudo Kazuo Ishiguro Kanehara Hitomi Dazai Osamu” AI翻訳を検証してみた：フィッツジェラルド、カズオ・イシグロ、金原ひとみ、太宰治” [Testing AI translation: Fitzgerald, Kazuo Ishiguro, Hitomi Kanehara, and Osamu Dazai].” note, September 7, 2023. Available at: https://note.com/happano/n/nb381c9b41955#0359461d-8149-4a4b-b8ce-7070f2ee5a7e
Kanehara, Hitomi. Autofiction. Translated by David Karashima. London: Vintage Books, 2007.
Kinsui, Satoshi. Bācharu Nihongo yakuwarigo no nazo ヴァーチャル日本語：役割語の謎 [Virtual Japanese: Enigmas of Role Language]. Tokyo: Iwanami Shoten, 2003.
Lakoff, Robin. Language and Women’s Place. Cambridge: Cambridge University Press, 1973.
Mills, Sara. Gender and Politeness. Cambridge: Cambridge University Press, 2003.
Nakamura, Momoko. “Gengo ideorogī to shite no onna kotoba” 言語イデオロギーとしての「女ことば」[Linguistic Ideology of Japanese Women’s Language]. Studies in Pragmatics 7 (2005):109-122.
Nakamura, Momoko. Hon’yaku ga tsukuru nihongo niroin wa onna kotoba o hanashi tsuzukeru 翻訳がつくる日本語：ヒロインは「女ことば」を話し続ける [Japanese Formed by Translation: Heroines Keep Using ‘Onna Kotoba’]. Tokyo: Hakutakusha, 2013.
Okamoto, Shigeko and Satō, Shie. “Less Feminine Speech among Young Japanese Females.” In Locating Power: Proceedings of the Second Berkeley Women and Language Conference, Vol.1, edited by Kira Hall et al. Berkeley and Calif: Berkeley Women and Language Group, 1992.
Okamoto, Shigeko. “Wakai josei no otoko kotoba: Kotoba zukai to aidentiti” 若い女性の「男ことば」：言葉づかいとアイデンティティ [Young women’s otoko kotoba: language use and identities]. In Jendā de manabu gengogaku ジェンダーで学ぶ言語学” [Learning linguistics through gender], edited by Nakamura Momoko. Sekai Shisōsha, 2010.
Prates, Marcelo, Avelar, Pedro, and Lamb, Luis C. “Assessing Gender Bias in Machine Translation: A Case Study with Google Translate.” Neural Computing and Applications 32, no .1 (2020): 6363-6381.
Reiss, Katharina. “Text Types, Translation Types and Translation Assessment.” In Readings in translation theory, 1977/89.
Tanaka, Lidia. “Language, Gender, and Culture.” In The Routledge Handbook of Language and Culture, edited by Farzad Sharifian. London: Routledge, 2014.
Usami, Mayumi. “Ibunka sesshoku to poraitonesu: Disukōsu poraitonesu no kanten kara” 異文化接触とポライトネス: ディスコース・ポライトネス理論の観点から [Intercultural Interactions and Politeness: from the Standpoint of the Discourse-Politeness Theory].” Studies in the Japanese Language 54, no 3 (2003): 117-132.
Usami, Mayumi. “On the Notion of ‘Discourse Politeness’: Based on the Analysis of Japanese Conversations”. Paper presented at the International Symposium on Linguistic Politeness: Theoretical Approaches and Intercultural Perspectives, Chulalongkorn University, 1999.
Vermeer, Hans J. “Skopos and Commission in Translational Action.” In Readings in Translation Theory, edited by Andrew Chesterman. Helsinki: Oy Finn Lectura Ab, 1989/2004.
Cover Image: Photo by appshunter.io on Unsplash.
