Cultural Fidelity in English-to-Hindi Translation: A Preservation-Fluency Frontier for Gender Recoverability

Researchers developed methods to preserve gender information in English-to-Hindi machine translation, a challenge caused by Hindi's ergative and honorific grammatical structures. Two inference-time interventions—Source-Aware Reranker and Phenomenon-Aware Reranker—significantly improved gender preservation but revealed a tradeoff between cultural fidelity and translation fluency.

This research addresses a fundamental problem in neural machine translation: how generative systems handle culturally and linguistically specific features that don't map cleanly across language pairs. Translation isn't merely linguistic substitution—it's a process that shapes how cultural meaning travels globally. When gender information present in English source text disappears in Hindi output, downstream users lose critical social context about persons, relationships, and identities.

The study's scale and methodology matter. Testing 37,345 instances across twelve categories provides robust empirical evidence that this gender erasure is systematic rather than anecdotal. The researchers identified specific grammatical mechanisms—ergative constructions and honorifics—that obscure gender, enabling targeted interventions rather than black-box solutions.

The results present a genuine dilemma for practitioners. Phenomenon-Aware Reranking achieved dramatic improvements in gender preservation (10.3% to 81.3% in human evaluation) but degraded fluency from 4.36 to 3.37 on a 5-point scale. This isn't a technical failure to be solved but a legitimate preservation-fluency frontier where gains in one dimension require sacrifices in another. For applications prioritizing cultural accuracy—legal documents, medical communications, identity-related content—this tradeoff favors preservation. For content emphasizing readability, the current systems may suffice.

The work signals growing maturity in culturally-aware AI development. Rather than assuming universal translation principles, researchers now examine how specific language pairs require mechanism-aware solutions. This approach has implications for other language pairs with grammatical features misaligned with English, and broader patterns in localization where cultural context matters as much as linguistic accuracy.

• →Machine translation systems systematically erase gender information when translating English to Hindi due to grammatical structures in the target language.
• →Phenomenon-Aware Reranking improved gender preservation from 10.3% to 81.3% but reduced fluency from 4.36 to 3.37, demonstrating an unavoidable preservation-fluency tradeoff.
• →Cultural fidelity in translation requires mechanism-aware interventions that target specific grammatical phenomena rather than general approach improvements.
• →Different use cases demand different optimization priorities: legal and identity documents may prioritize preservation while general content prioritizes fluency.
• →This research exemplifies how AI localization for non-English languages requires language-pair-specific solutions rather than universal translation algorithms.