PAT: RAG-Based LLM Translation Beyond Sentence-Level MT

PAT tests whether retrieval, user specifications, and whole-document LLM generation can move automatic translation beyond sentence-by-sentence transfer.

TL;DR — PAT is a RAG-based system for whole-document, corpus-informed draft translation. It combines user-configured specifications with paragraph-, section-, and document-level examples from a comparable U.S. English and Latin American Spanish corpus. In an evaluation of six translations, a limited prompt produced no meaningful reformulation, while specification- and corpus-informed translations sometimes produced substantial but uneven reformulation.

Problem and research question

The abstract frames the problem directly: automatic translation systems, including CAT tools and MT, "overwhelmingly treat translation as a sentence-by-sentence act." The paper asks whether LLMs can be moved beyond that paradigm through "whole-document, corpus-informed translation."

The supported research question is therefore not simply whether an LLM can translate. It is whether an LLM-based system can be steered toward document-level reformulation that fits the target Spanish-language context.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.