[Paper Review] Gender Neutralization for an Inclusive Machine Translation: from Theoretical Foundations to Open Challenges

Andrea Piergentili, Dennis Fucci+3 more|arXiv (Cornell University)|Jan 24, 2023

Gender Studies in Language11 citations

TL;DR

The paper defines gender-neutral translation (GNT) for English→Italian MT, reviews guidelines on gender-inclusive language, and outlines desiderata, challenges, and potential solutions for implementing GNT in MT.

ABSTRACT

Gender inclusivity in language technologies has become a prominent research topic. In this study, we explore gender-neutral translation (GNT) as a form of gender inclusivity and a goal to be achieved by machine translation (MT) models, which have been found to perpetuate gender bias and discrimination. Specifically, we focus on translation from English into Italian, a language pair representative of salient gender-related linguistic transfer problems. To define GNT, we review a selection of relevant institutional guidelines for gender-inclusive language, discuss its scenarios of use, and examine the technical challenges of performing GNT in MT, concluding with a discussion of potential solutions to encourage advancements toward greater inclusivity in MT.

Motivation & Objective

• Define gender-neutral translation (GNT) as translating without marking gender in the target language when not inferable from the source.
• Survey institutional guidelines in English and Italian to understand conceptualizations of gender and neutrality.
• Outline practical desiderata (D1-D3) for applying GNT in MT.
• Discuss technical challenges in context detection, data, and evaluation for GNT in MT.
• Propose potential methodological directions to advance inclusive MT research.

Proposed method

• Review and synthesize gender and language theory to identify how gender is encoded across languages and where biases arise in MT.
• Map and compare 30 institutional guidelines in English and Italian to extract neutralization strategies and their applicability.
• Define a three-desiderata framework (D1-D3) for when and how to apply GNT in MT with concrete examples.
• Discuss dynamic, data, and evaluation challenges for implementing GNT in MT and review possible technical approaches (context, constraints, decoding, and evaluation).
• Provide a roadmap outlining future research directions and potential solutions for GNT in MT.

Experimental results

Research questions

• RQ1 What constitutes gender-inclusive language and how can it be operationalized as GNT in MT?
• RQ2 When should MT systems apply gender-neutral translations, and how to detect gender cues in source texts?
• RQ3 What neutralization strategies from English and Italian guidelines are transferable to MT?
• RQ4 What are the main technical challenges in data, decoding, and evaluation for GNT?
• RQ5 What roadmaps and methodologies can advance inclusive MT toward robust GNT?

Key findings

• Gender-neutral translation is defined as automatically translating into the target language without marking the gender of human referents when not inferable from the source.
• English guidelines focus on neutralizing pronouns and gender cues, while Italian guidelines emphasize neutralizing gendered nouns, highlighting cross-language differences.
• GNT requires careful consideration of trade-offs between neutrality, fluency, and context, especially in formal domains where neutrality may be constrained by genre.
• GNT is a dynamic constraint: decisions on applying neutralization depend on source gender cues and discourse context, potentially benefiting from document-level context rather than sentence-level translation.
• There is a data scarcity challenge: training MT models with neutral target forms is difficult due to lack of parallel data with gender-neutral translations.
• Constraint-based decoding and re-ranking approaches offer potential paths to enforce GNT, but may impact translation quality and require careful evaluation.
• There is a need for dedicated evaluation protocols and benchmarks to assess gender neutrality separately from overall translation quality, including datasets like parallel source-typed and neutralized targets.