Enhancing Gender-Inclusive Machine Translation with Neomorphemes and Large Language Models

Overview

• Machine translation (MT) models are known to have gender bias, especially when translating into languages with extensive gendered morphology.
• They often fail to use gender-inclusive language that represents non-binary identities.
• This paper explores the use of gender-inclusive "neomorphemes" - new linguistic elements that avoid binary gender markings - as an approach to fairer MT.
• The researchers investigate prompting techniques with large language models (LLMs) to translate from English to Italian using neomorphemes.
• They release a new resource called Neo-GATE to evaluate gender-inclusive En-It translation with neomorphemes, and assess four LLMs on this novel MT task.

Plain English Explanation

Machine translation (MT) models, which convert text from one language to another, can often have biases when it comes to gender. This is especially true when translating into languages that have extensive gendered grammar, like verb endings or pronouns that change based on the gender of the subject. These MT models sometimes fail to use inclusive language that represents non-binary identities.

To address this issue, the researchers in this paper looked at using "neomorphemes" - new linguistic elements that avoid binary male/female gender markings. They explored techniques for prompting large language models (LLMs), which are powerful AI models trained on vast amounts of text, to translate from English to Italian using these neomorphemes.

Since this is a relatively new area, the researchers also created a new evaluation resource called Neo-GATE. This allows them to assess how well different LLM models perform at this gender-inclusive translation task using neomorphemes.

The key idea is to move beyond the traditional male/female gender binary that most MT systems are designed for, and find ways to generate translation outputs that are more representative of the full spectrum of gender identities.

Technical Explanation

The paper investigates the use of gender-inclusive "neomorphemes" as an approach to address gender bias in machine translation (MT) models, especially when translating into languages with extensive gendered morphology like Italian.

The researchers explore prompting techniques with large language models (LLMs) to translate from English to Italian using these neomorphemes. They assess four different LLM families and sizes, as well as various prompt formats, to identify the strengths and weaknesses of each on this novel MT task.

To enable this evaluation, the researchers release a new resource called Neo-GATE, designed to assess gender-inclusive En-It translation with neomorphemes. This helps fill an important gap, as this area has been under-explored due to its novelty and lack of publicly available evaluation data.

The experiments show that while current LLMs can generate some gender-inclusive translations using neomorphemes, there is still significant room for improvement. The models often struggle with consistency, morphological agreement, and accurately representing non-binary identities. The paper discusses the implications of these findings and highlights directions for future research in this emerging area of fairer machine translation.

Critical Analysis

The paper makes a valuable contribution by exploring the use of gender-inclusive neomorphemes as an approach to address bias in machine translation. This is an important issue, as MT models that fail to represent the full spectrum of gender identities can perpetuate harmful stereotypes and marginalization.

However, the research is still quite preliminary, as the authors acknowledge. The evaluation is limited to a single language pair (En-It), and the neomorpheme vocabulary used is relatively small. There may be challenges in scaling this approach to broader use cases and language contexts.

Additionally, the paper does not delve deeply into the societal and ethical implications of this work. While the intent is to promote more inclusive translation, the real-world impact on marginalized communities is not fully explored. Further research is needed to understand how these techniques would perform in high-stakes applications.

Despite these limitations, the paper lays important groundwork for future developments in this area. The Neo-GATE resource in particular is a valuable contribution that can spur more research and innovation in gender-inclusive machine translation. As the field continues to evolve, it will be critical to center the perspectives of diverse stakeholders and ensure these technologies empower rather than further marginalize underrepresented groups.

Conclusion

This paper presents a novel approach to addressing gender bias in machine translation by exploring the use of gender-inclusive neomorphemes. Through the release of the Neo-GATE evaluation resource and assessment of different large language models, the researchers shed light on both the potential and limitations of this technique.

While current MT models struggle to consistently and accurately represent non-binary identities using neomorphemes, this work represents an important step towards fairer and more inclusive machine translation. As language technologies become increasingly ubiquitous, it is crucial that they evolve to better reflect the full diversity of human experience.

The insights from this paper can inform future research directions in this emerging area, ultimately working towards MT systems that empower rather than marginalize underrepresented groups. Continued advancements in this direction, combined with thoughtful consideration of the societal implications, hold promise for more equitable and representative language technologies.