💬 NLP

G^2C-MT: Graph-Guided Context Selection for Document-Level Machine Translation

The paper proposes G^2C-MT, a novel framework that models document-level machine translation context selection as a structured path discovery problem on a lightweight discourse graph using depth-biased random walks to effectively capture long-range dependencies and outperform existing baselines across various large language models.

Original authors: Baijun Ji, Zixuan Zhou, Xiangyu Duan, Yu Liu, Longbo Sun, Rupu Wei, Bohong Zhao

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Baijun Ji, Zixuan Zhou, Xiangyu Duan, Yu Liu, Longbo Sun, Rupu Wei, Bohong Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to translate a long, complex novel from English to Chinese. If you only look at one sentence at a time, you might miss the bigger picture. For example, a character might be referred to as "he" in sentence 50, but you won't know who "he" is until you read sentence 5. Or, a specific technical term might be defined in the first chapter but used again in the last chapter. If you forget that definition, your translation will be wrong.

This is the problem G2C-MT (Graph-Guided Context for Machine Translation) tries to solve. It's a new way to help Artificial Intelligence (AI) translate entire documents, not just single sentences.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Amnesia" and "Noise" Issues

Current AI translators often struggle with two things:

The Window Problem: Some translators only look at the last few sentences (like looking through a small window). If the answer to a question is 50 sentences back, they miss it.
The "Bag of Words" Problem: Other translators try to find any sentence that sounds similar to the current one. But this is like finding a sentence about "apple" because the current sentence is about "apple pie," even if the "apple" in the current sentence is actually a computer brand. They grab the wrong context, creating confusion.

2. The Solution: Building a "Map" of the Story

Instead of just looking at the last few sentences or grabbing random similar ones, G2C-MT builds a map (a graph) of the entire document before it starts translating.

Think of the document as a city, where every paragraph is a building.

The Roads (Edges): The system draws roads between buildings. But it doesn't just draw roads between neighbors (like building 1 and building 2). It also draws roads between buildings that share:
- Similar Themes: (Semantic similarity)
- Shared Keywords: (Like a specific name or technical term appearing in both)
- Order: (Which building comes right after the other)

This creates a Discourse Graph. It's a lightweight map that shows not just who is next to whom, but who is connected to whom across the whole document.

3. The Journey: The "Depth-Biased Hiker"

Once the map is built, the AI needs to decide which past paragraphs to read before translating the current one.

Instead of just walking to the next door (like a standard window) or jumping to the most similar building (like a random search), G2C-MT uses a hiker with a special compass.

The Hiker's Goal: The hiker starts at the current building and walks backward through the map to find the most helpful history.
The Compass (Depth Bias): The hiker is biased to walk deeper into the past. If a building has a long chain of connections leading back to the beginning of the story, the hiker is more likely to follow that path. This ensures the AI finds the "root" of a story or a definition, even if it's far away.
The Path: The hiker traces a single, logical path of connected paragraphs. This path becomes the "context" the AI uses to translate the current sentence.

4. The Safety Net: Trying Multiple Paths

Sometimes, a story is ambiguous. There might be two different logical ways to connect the dots.

G2C-MT doesn't just take one path. It sends out multiple hikers at the same time, each taking a slightly different route through the map.
Each hiker brings back a different translation suggestion.
The system then looks at all the suggestions and picks the one that makes the most sense (the "majority vote" or the most consistent one). This makes the translation more robust and less likely to make a silly mistake.

5. Why It's Better (The Results)

The authors tested this on two types of documents:

Technical Manuals (SAP): These are dry, step-by-step instructions where terms must be consistent.
Narrative Stories (TED Talks): These have loose structures and long-range themes.

The Findings:

Better Accuracy: G2C-MT consistently beat other methods. It translated technical terms correctly because it could "remember" the definition from 10 paragraphs ago.
Better Flow: It kept the story coherent because it followed the logical path of the text, not just random similar words.
Efficiency: Unlike other methods that require the AI to "think" about the whole document structure using expensive, slow processes, G2C-MT builds its map quickly using simple math (like counting keywords and measuring similarity) and then walks the map. It's fast and doesn't waste computer power.

Summary Analogy

Imagine you are translating a mystery novel.

Old Way (Window): You only read the last 3 pages. You miss the clue from page 1.
Old Way (Random Search): You search for the word "murder" and find a page from a completely different book that also mentions "murder," confusing your translation.
G2C-MT: You have a smart detective who draws a map of the whole book. When you get to a confusing clue, the detective walks backward along the map, following the specific trail of clues that leads back to the beginning, finds the original definition, and tells you exactly what it means. If the trail is tricky, the detective sends out a team to check a few different routes and picks the one that solves the mystery best.

This paper claims that this "map-based" approach helps AI translate long documents with much higher quality and consistency than previous methods.

Technical Summary: G2C-MT

Problem Statement
Document-level machine translation (DocMT) requires capturing long-range discourse dependencies to ensure lexical consistency, coreference resolution, and logical coherence. While Large Language Models (LLMs) have shown promise in handling long contexts, translating entire documents in a single pass often leads to sentence omissions, context dilution, and prohibitive decoding costs due to the quadratic complexity of LLMs relative to input length. Existing solutions attempt to mitigate this via retrieval-based methods or graph-based strategies. However, retrieval-based approaches often treat sentences as an unstructured collection, failing to preserve explicit discourse structures. Conversely, existing graph-based methods frequently rely on expensive LLM-based relation classification for edge definitions and are limited to selecting first-order neighbors, preventing the capture of deep, multi-hop discourse paths essential for global document structure.

Methodology
The authors propose G2C-MT (Graph-Guided Context for Machine Translation), a framework that reframes context selection as a structured path discovery problem on a lightweight discourse graph. The pipeline consists of three stages:

Directed Discourse Graph Construction:
The source document is modeled as a weighted directed acyclic graph (DAG) where each paragraph is a node. Directed edges connect a target paragraph to previous paragraphs ( $j < i$ ), reflecting the sequential nature of translation. Edge weights ( $w_{ij}$ ) are calculated via a fusion of three factors:
- Semantic Relevance ( $S_{sem}$ ): Cosine similarity between paragraph embeddings, truncated by a threshold to reduce noise.
- Sequential Adjacency ( $S_{seq}$ ): A fixed weight for immediately adjacent paragraphs to ensure local coherence.
- Keyword Overlap ( $S_{key}$ ): A TF-IDF based score to capture term consistency (e.g., proper nouns) that semantic similarity might miss.
  This construction is lightweight, relying on pre-trained embeddings and TF-IDF rather than expensive LLM inference.
Depth-Biased Context Path Sampling:
To select historical context for a target paragraph, the method employs a depth-biased random walk starting from the target node and backtracking.
- Depth Heuristic: A dynamic programming approach calculates $\phi(v)$ , representing the longest path length back from a node, estimating its potential to provide rich historical context.
- Probabilistic Transition: The walker transitions to previous nodes based on a probability distribution that combines the edge weight ( $w_{ij}$ ) with a bias term $(\phi(v_{next}) + 1)^\lambda$ . The hyperparameter $\lambda$ controls the preference for deeper, more informative context paths over immediate neighbors.
- Multi-Path Sampling: Due to the stochastic nature of the walk, multiple distinct context paths can be sampled for a single target.
Path-Based Contextual Generation:
The sampled path is reversed to restore natural document order and formatted into a structured prompt containing the source paragraphs and their corresponding translations. This prompt is fed to an LLM to generate the translation for the current target paragraph.
- Robustness via Aggregation: For ambiguous inputs, the framework can generate multiple translations from different sampled paths. The final output is selected by clustering these candidates (via k-means) and choosing the one closest to the cluster centroid.

Key Contributions

Lightweight Graph Modeling: The paper introduces a novel framework that models document context as a DAG using a computationally efficient construction process (embedding + TF-IDF) rather than expensive LLM-based relation classification.
Depth-Biased Traversal: A biased random walk mechanism is designed to explicitly favor deeper context paths, enabling the retrieval of long-range, non-linear dependencies that fixed-window or first-order neighbor methods miss.
Multi-Path Robustness: The probabilistic traversal naturally supports multi-path sampling, allowing the aggregation of diverse translation candidates to improve robustness against discourse-level ambiguity.

Experimental Results
The method was evaluated on two benchmarks: the SAP dataset (technical IT domain) and the IWSLT 2017 dataset (narrative/TED talks), using various LLM backbones (DeepSeek-V3, Gemini-2.5-Flash-lite, Qwen-2.5/3 series).

Performance: G2C-MT consistently outperformed strong baselines, including Sentence-level, Window-Based, Semantic-Based retrieval, and agentic frameworks like GRAFT and DelTA. It achieved the highest d-BLEU scores across all translation directions and model backbones.
Discourse Metrics: On discourse-specific metrics (BlonDe, d-Prism, d-Comet), G2C-MT achieved the best scores, demonstrating superior handling of entities, tenses, and pronouns compared to methods that rely solely on local windows or global semantic similarity.
Efficiency: Unlike agent-based methods requiring multiple LLM calls for context curation, G2C-MT requires exactly $N$ LLM calls for $N$ paragraphs, as graph construction and path sampling are pre-processing steps.
Ablation: Removing the keyword edge or the depth-biased random walk resulted in the largest performance drops, highlighting the importance of terminological consistency and deep context retrieval.

Significance and Claims
The paper claims that G2C-MT provides a robust and computationally efficient solution for document-level translation by explicitly modeling structured discourse dependencies. By viewing context selection as a path discovery problem on a graph, the method successfully bridges the gap between local coherence (captured by sequential edges) and global relevance (captured by semantic and keyword edges). The authors assert that their approach avoids the trade-off between fixed-window limitations and the noise of unstructured semantic retrieval, offering a scalable alternative to complex multi-agent frameworks while maintaining high translation quality and coherence.