TLDR: TRANSGRAPH is a novel framework for document-level machine translation that uses structured discourse graphs to model inter-chunk relationships. Instead of relying on exhaustive or sequential context, it selectively conditions LLM-based translation on relevant graph neighborhoods. This two-stage process involves LLM-guided document chunking and discourse graph construction, followed by iterative chunk translation using graph-retrieved, relation-labeled context. TRANSGRAPH consistently outperforms existing baselines in translation quality and terminology consistency while significantly reducing computational costs and improving discourse cohesion.
Large Language Models (LLMs) have brought about a revolution in machine translation, showing impressive capabilities at the sentence level. However, translating entire documents, especially long ones, still presents significant challenges. The main hurdles include maintaining consistent meaning over long stretches of text and ensuring that different parts of the document relate to each other coherently, a concept known as discourse coherence. Traditional methods struggle with these long-range dependencies, and even newer ‘agentic’ machine translation systems, which use multiple LLM agents and memory, often demand substantial computing power and can be sensitive to how they retrieve information.
Addressing these challenges, researchers Viet-Thanh Pham, Minghan Wang, Hao-Han Liao, and Thuy-Trang Vu from Monash University have introduced a novel framework called TRANSGRAPH. This system offers a discourse-guided approach to document translation, moving away from simply relying on sequential text or exhaustive context. Instead, TRANSGRAPH explicitly models the relationships between different sections of a document using structured ‘discourse graphs’. This allows the translation process to selectively focus on only the most relevant contextual information for each segment, rather than processing the entire document or a large, undifferentiated chunk of text.
How TRANSGRAPH Works: A Two-Stage Process
TRANSGRAPH operates in two main stages to achieve its goal of coherent and accurate document translation:
Stage 1: Document Preparation
First, the source document is intelligently broken down into smaller, coherent textual units called ‘chunks’. This chunking process is performed by an LLM, ensuring that sentence boundaries are preserved and that each chunk contains related sentences. This step is crucial for managing the context window limitations of LLMs. Once chunked, TRANSGRAPH constructs a directed, labeled graph. In this graph, each chunk is a node, and the connections (edges) between chunks represent specific discourse relations. These relations, such as ‘Entity-Coreference’ (when two chunks refer to the same entity) or ‘Motivation→Method’ (when one chunk describes a problem and another its solution), are identified by an LLM based on established discourse theories. This graph explicitly maps out how different parts of the document are logically connected.
Stage 2: Iterative Chunk Translation
With the discourse graph in place, translation proceeds chunk by chunk. For each chunk to be translated, TRANSGRAPH doesn’t look at all preceding text. Instead, it consults the discourse graph to identify its ‘in-neighbors’ – those chunks that provide direct, relevant context based on the established discourse relations. A compact ‘context package’ is then assembled, containing the content of these neighboring chunks along with their specific relation labels. This structured context is then fed to an LLM, which translates the current chunk. By conditioning the translation on these precise, relation-labeled contexts, the system significantly improves the consistency of terminology, resolves references more accurately, and maintains the overall rhetorical coherence of the document. After all chunks are translated, they are reassembled in their original order to form the complete target document.
Also Read:
- TeaRAG: Enhancing Language Models with Efficient Retrieval and Reasoning
- Decoding Contracts: How AI Transforms Legal Document Analysis into Visual Graphs
Key Advantages and Performance
TRANSGRAPH offers several significant advantages. It selectively conditions translation on only the contextual information that truly matters, leading to more efficient processing. It consistently improves terminology consistency and overall cohesion with a much lower ‘token overhead’ compared to other advanced systems. Furthermore, the framework is flexible, meaning it can work with various underlying LLM architectures.
Experiments conducted on three diverse document-level machine translation benchmarks, spanning six languages and different domains (including technical presentations and literary texts), demonstrated TRANSGRAPH’s superior performance. It consistently outperformed traditional sentence-level and single-pass document translation methods, as well as more complex agentic MT systems, in terms of translation quality (measured by BLEU and COMET scores) and, notably, terminology accuracy. For instance, on terminology-focused benchmarks like BWB and ACL 60/60, TRANSGRAPH achieved significantly higher accuracy in translating specialized terms.
A detailed cost analysis revealed that TRANSGRAPH is substantially more efficient than multi-agent frameworks, requiring one order of magnitude fewer tokens and calls to LLMs, while still delivering stronger translation performance. Cohesion analysis, using an LLM-as-a-Judge approach, confirmed that TRANSGRAPH significantly enhances the handling of coreference (ensuring consistent references to entities) and conjunctions (preserving logical relationships between sentences).
While TRANSGRAPH represents a major step forward, the authors acknowledge limitations such as reliance on LLM accuracy for relation labeling and potential scalability challenges for extremely long documents with very dense relation structures. Future work aims to further improve robustness and efficiency through techniques like retrieval-augmented compression.
This research highlights that a modest, well-structured approach to discourse retrieval can be a powerful tool for enhancing document translation with large language models. You can read the full research paper here.
