Can an Old Dog Be Taught New Tricks? Taking LLMs Beyond Sentence Level Translation

TL;DR

The paper introduces PAT, a RAG-based system that enables large language models to generate translations at the document level by leveraging comparable corpora and user-defined specifications. It evaluates whether such systems can produce reformulated translations that adapt to target-language discourse and rhetorical norms, moving beyond traditional sentence-by-sentence approaches. Results show that while some reformulation is possible, effectiveness remains inconsistent.

Why It Matters

Professional translation workflows often require more than literal word-for-word conversion; adapting tone, structure, and cultural context is essential. This work demonstrates a path toward more context-aware, document-level translation using LLMs, which could improve quality in real-world applications like legal, academic, and technical translation.

Key Findings

• Using a retrieval-augmented generation approach with a comparable corpus enables LLMs to generate document-level translations that reflect target-language discourse patterns.
• User-configured specifications significantly influence the degree of reformulation, but their impact varies depending on the content and context.
• While some translations showed meaningful adaptation to target-language norms, the overall effectiveness of reformulation was inconsistent, indicating room for improvement in prompt design and retrieval quality.