Discourse Graph Guided Document Translation with Large Language Models

TL;DR

TransGraph is a novel discourse-guided framework for document translation that models inter-chunk relationships with structured graphs, improving translation quality and efficiency over existing methods.

Contribution

It introduces a discourse graph-based approach to explicitly model long-range dependencies in document translation, reducing computational overhead.

Findings

Outperforms strong baselines in translation quality.

Enhances terminology consistency across documents.

Reduces token overhead compared to existing methods.

Abstract

Adapting large language models to full document translation remains challenging due to the difficulty of capturing long-range dependencies and preserving discourse coherence throughout extended texts. While recent agentic machine translation systems mitigate context window constraints through multi-agent orchestration and persistent memory, they require substantial computational resources and are sensitive to memory retrieval strategies. We introduce TransGraph, a discourse-guided framework that explicitly models inter-chunk relationships through structured discourse graphs and selectively conditions each translation segment on relevant graph neighbourhoods rather than relying on sequential or exhaustive context. Across three document-level MT benchmarks spanning six languages and diverse domains, TransGraph consistently surpasses strong baselines in translation quality and terminology consistency while incurring significantly lower token overhead.