Source-primed Multi-turn Conversation Helps Large Language Models Translate Documents

TL;DR

The paper tackles the challenge of document-level machine translation with large language models by addressing omission errors through a simple, training-free multi-turn conversational framework. It translates documents in turns over segments while reusing previous context via a KV-cache, and introduces a source-primed variant that exposes the full source upfront to improve coherence from the start. Across multiple decoder-only instruction models and the WMT-24 General Track, the proposed method consistently outperforms single-turn and segment-level baselines on BLEU, COMET-22, and BlonDE metrics, with the source-primed version delivering the strongest results. The approach demonstrates that document-level translation can be effectively achieved without additional training, offering a scalable, efficient path for improving translation quality in chat-enabled LLMs while highlighting domain-specific benefits and limitations.

Abstract

LLMs have paved the way for truly simple document-level machine translation, but challenges such as omission errors remain. In this paper, we study a simple method for handling document-level machine translation, by leveraging previous contexts in a multi-turn conversational manner. Specifically, by decomposing documents into segments and iteratively translating them while maintaining previous turns, this method ensures coherent translations without additional training, and can fully re-use the KV cache of previous turns thus minimizing computational overhead. We further propose a `source-primed' method that first provides the whole source document before multi-turn translation. We empirically show this multi-turn method outperforms both translating entire documents in a single turn and translating each segment independently according to multiple automatic metrics in representative LLMs, establishing a strong baseline for document-level translation using LLMs.

Figures (2)

• Figure 1: Different settings of document-level translation using LLMs.
• Figure 2: Token number across top N longest docs