Gender-Neutral Machine Translation Strategies in Practice

Abstract: Gender-inclusive machine translation (MT) should preserve gender ambiguity in the source to avoid misgendering and representational harms. While gender ambiguity often occurs naturally in notional gender languages such as English, maintaining that gender neutrality in grammatical gender languages is a challenge. Here we assess the sensitivity of 21 MT systems to the need for gender neutrality in response to gender ambiguity in three translation directions of varying difficulty. The specific gender-neutral strategies that are observed in practice are categorized and discussed. Additionally, we examine the effect of binary gender stereotypes on the use of gender-neutral translation. In general, we report a disappointing absence of gender-neutral translations in response to gender ambiguity. However, we observe a small handful of MT systems that switch to gender neutral translation using specific strategies, depending on the target language.

• The paper evaluates 21 machine translation systems using a controlled test suite to assess their ability to maintain gender neutrality and ambiguity when translating from English to languages like Spanish, Icelandic, and Czech.
• A key finding is the prevalent masculine default observed across most systems when faced with gender ambiguity, and the limited adoption of explicit gender-neutral strategies.
• The findings highlight significant gaps in existing systems regarding gender-neutral translations, suggesting future research directions including the use of large language models and refined test sets for improvement.

Gender-Neutral Machine Translation Strategies in Practice

The paper "Gender-Neutral Machine Translation Strategies in Practice" addresses the critical and challenging task of preserving gender neutrality in machine translation (MT) systems, specifically when translating from English, a notional gender language, to grammatical gender languages such as Spanish, Icelandic, and Czech. It examines 21 MT systems to ascertain their capacity to maintain gender ambiguity from source material and discusses the implementation of gender-neutral strategies in translation.

The authors highlight the importance of this research given the increasing relevance of gender-neutral language in linguistic communities and the potential harms of misgendering, especially in the context of non-binary individuals. Misgendering due to gender ambiguity in source texts often results from stereotypes or the prevalent default usage of masculine forms. The paper focuses on scenarios where gender cannot be determined from the source text, which is common in English, and evaluates if this ambiguity is appropriately transferred to the target language.

Study Setup and Observations

A detailed study methodology is utilized, involving a test suite specifically designed for this research, where source sentences are presented in gender-determined and gender-ambiguous pairs. This setup allows for controlled measurement of translation sensitivity to gender neutrality.

Key observations from the study include:

1. Baseline Gender Neutrality: The initial frequency of gender-neutral translations varies significantly by language due to the natural availability of gender-neutral lexicon. Icelandic and Czech exhibit lower baseline neutrality rates compared to Spanish.
2. Masculine Default Response: A prevalent masculine default is observed across systems when faced with gender ambiguity, indicating a significant area for improvement in achieving gender-neutral translations.
3. Gender-Neutral Response: The paper finds limited adoption of gender-neutral strategies across MT systems. Notably, only a few systems demonstrate an increased use of gender-neutral translations in response to active gender ambiguity, such as using "they" in English. The specific strategies employed also vary, with some systems utilizing alternative morphological forms or gender-neutral synonyms.

Gender Stereotypes and Translation

The influence of gender stereotypes on translation outcomes is thoroughly investigated. While stereotypes impact the gendered nature of translations, they do not appear to affect gender-neutral translations. This suggests that the base choice of adjective overrides subsequent gender agreement considerations, reinforcing the challenge for MT systems to switch strategies based on source text requirements.

Implications and Future Research

The findings highlight significant gaps in existing MT systems concerning gender-neutral translations, despite available strategies in the target languages. Future research avenues are suggested, including the exploration of LLMs combined with innovative prompting techniques as a promising approach to improving gender-sensitive translations. However, attention must be paid to the resultant translation quality and coherence, emphasizing the need for a balanced approach between baseline gender neutrality and gender-neutral response.

Moreover, the limited success of some systems in adopting gender-neutral strategies underlines the necessity of further investigation into system-specific characteristics that facilitate such translations. The authors propose refining and expanding the controlled template-based test sets to enhance training data for more complex real-world utilization scenarios.

Conclusion

This paper sheds light on the sophisticated challenge of gender-neutral translation in machine translation systems, a crucial task in advancing towards more inclusive language technologies. While the study finds significant room for improvement, certain approaches, notably using LLMs, offer potential pathways for development. Continuous evaluation and enhancement of gender-neutral strategies are essential to mitigate representational harms and achieve greater inclusivity in machine translation.