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

TL;DR — This paper introduces PAT (Pragmatic Auto-Translator), a RAG-based system that moves LLM translation beyond the sentence-by-sentence paradigm toward whole-document, corpus-informed generation. It matters because it targets pragmatic/discourse-level reformulation—not just literal accuracy—for professional translation drafts.

• System design: PAT pairs user-configured specifications with retrieved paragraph-, section-, and document-level examples from a comparable corpus of authentic longform U.S. English and Latin American Spanish texts, feeding these to an LLM for whole-document translation.
• Evaluation: Six automatic translations of essays on generative AI across three projects were assessed by two trained evaluators (U.S. English → LATAM/Mexican Spanish) using a customized MQM typology.
• Findings: Limited prompts produced no meaningful reformulation, while specification- and corpus-informed prompts sometimes yielded substantial reformulation—though "not always to effect"; LLMs can be steered toward reformulation, but effectiveness needs improvement.
• Contributions: The authors also discuss considerations for translation system design, comparable-corpus construction, and quality-evaluation methodology.