Title: Moving Past Generic Translations: Aligning Machine Translation with Audience and Intent

Abstract:

The quality of a translation is inherently tied to its intended purpose; identical source texts require varying translations based on the target audience, desired tone, and communicative goal. However, current machine translation (MT) systems and evaluation metrics typically view translation as a static, fixed mapping from source to target language. While Large Language Models (LLMs) allow users to explicitly define intent alongside the source text, this feature has yet to be assessed on a large scale.

This study presents a comprehensive evaluation of purpose-driven MT, spanning 50 languages, eight text domains, and eight different model sizes. Our findings reveal four key insights:

1. Impact of Explicit Instructions: Providing clear instructions significantly enhances the adaptability of translations. These improvements are particularly pronounced in informal domains such as social media and conversation, in larger model sizes, and within higher-resource languages.
2. Superiority Over Contextual Examples: Explicit instructions prove more effective than semantically similar few-shot examples or paragraph-level context.
3. Limitations of Current Metrics: Standard MT metrics are ill-equipped to assess adaptation quality, frequently penalizing translations that are well-suited to their specific intent.
4. Self-Generated Instructions: In the absence of curated instructions, models can autonomously generate relevant directives from the surrounding document context, thereby recovering up to 80% of the adaptability advantage seen with hand-crafted instructions.

These results demonstrate that purpose-adapted MT is both a feasible and quantifiable capability of LLMs, underscoring the critical need for the development of metrics that account for communicative intent.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC