pith. sign in

arxiv: 2606.03078 · v1 · pith:E75DX5KVnew · submitted 2026-06-02 · 💻 cs.CL

G²C-MT: Graph-Guided Context Selection for Document-Level Machine Translation

Baijun Ji , Zixuan Zhou , Xiangyu Duan , Yu Liu , Longbo Sun , Rupu Wei , Bohong Zhao This is my paper

Pith reviewed 2026-06-28 10:34 UTC · model grok-4.3

classification 💻 cs.CL
keywords document-level machine translationcontext selectiondiscourse graphrandom walklarge language modelsDocMTgraph-guided selectionparagraph relations
0
0 comments X

The pith

Viewing document translation context selection as path discovery on a paragraph graph lets LLMs use structured discourse chains.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that document-level machine translation improves when context selection is treated as discovering paths on a graph of paragraphs rather than retrieving flat sets of sentences. Paragraphs become nodes whose connections reflect semantic similarity, document adjacency, and keyword overlap. A depth-biased random walk then traces backward paths from the target paragraph to supply context when prompting the LLM, with an option to sample multiple paths and combine results. Experiments across domains show gains over baselines when the same LLMs are used. A sympathetic reader would care because the method offers an explicit, lightweight way to handle long-range dependencies that often break coherence in document translation.

Core claim

G^2C-MT represents each paragraph as a node in a lightweight discourse graph whose edges encode semantic similarity, adjacency, and keyword overlap, then uses a depth-biased random walk to sample a backward context path for the target paragraph; the resulting path is inserted into the LLM prompt, and the framework optionally aggregates multiple such paths for robustness on ambiguous cases.

What carries the argument

The depth-biased random walk over the discourse graph, which samples structured backward context paths to insert into LLM prompts.

Load-bearing premise

A lightweight graph whose edges come only from semantic similarity, adjacency, and keyword overlap is enough to capture the structured long-range discourse relations that matter for translation.

What would settle it

An experiment in which the same LLMs receive prompts built from randomly chosen paragraphs rather than graph-walk paths and produce equal or higher translation quality on the same test sets.

Figures

Figures reproduced from arXiv: 2606.03078 by Baijun Ji, Bohong Zhao, Longbo Sun, Rupu Wei, Xiangyu Duan, Yu Liu, Zixuan Zhou.

Figure 1
Figure 1. Figure 1: Overview of the G2C-MT Framework. The process involves three stages: (1) Constructing a discourse graph considering semantic (α), sequential (β), and keyword (γ) cohesion; (2) Traversing a context path via Depth-Biased Random Walk (backtracking from target to history), where nodes with higher depth potential ϕ (e.g., x3) and edge score attract the walker; (3) Formatting the path into a structured prompt fo… view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of Context Distances. The red dashed line represents the hard cutoff of standard Window-based approach (Window=4). The blue histogram shows the context selected by G 2C-MT. Our method retains strong attention to local context (peak at ∆ < 10) while maintaining a long tail of retrieval capabilities extending up to 100 sentences back, capturing long-range depen￾dencies that linear models miss. 4… view at source ↗
read the original abstract

Effective document-level machine translation (DocMT) requires capturing long-range discourse dependencies. Recent work has explored retrieval-based and discourse-aware context selection. However, these approaches often lack an explicit mechanism for modeling structured discourse dependencies between distant paragraphs in a document. In this paper, we propose G^2C-MT (Graph-Guided Context for Machine Translation), which views DocMT context selection as a structured path discovery problem on a lightweight discourse graph, rather than retrieving unstructured context sets or relying on expensive LLM-based discourse modeling. In detail, we represent each paragraph as a node and model the relationship between each pair of nodes, considering their semantic similarity, adjacency, and keyword overlap. Furthermore, we propose a depth-biased random walk over the graph to sample a backward context path for each target paragraph. The context path will be used to prompt a large language model (LLM) for translation. This framework naturally supports multi-path context sampling, which can improve robustness by aggregating diverse translation candidates for discourse-ambiguous inputs. Experiments conducted across various domains show that G^2C-MT outperforms strong baselines on multiple LLMs, including DeepSeek-V3, Gemini-2.5-Flash-lite, and the Qwen-2.5/3 series.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes G²C-MT, a framework that formulates document-level MT context selection as structured path discovery on a lightweight discourse graph. Paragraphs are nodes; edges are defined by semantic similarity, adjacency, and keyword overlap. A depth-biased random walk samples backward context paths that are concatenated into LLM prompts; multi-path sampling is supported for robustness. Experiments across domains are reported to show consistent gains over strong baselines on DeepSeek-V3, Gemini-2.5-Flash-lite, and Qwen-2.5/3 models.

Significance. If the reported gains are reproducible and attributable to the graph-guided paths rather than prompt length or diversity, the approach supplies a lightweight, non-LLM discourse model that could reduce reliance on expensive retrieval or full-document prompting while still capturing long-range dependencies. The multi-path aggregation mechanism is a potentially useful robustness feature.

major comments (1)
  1. [Abstract] Abstract: the central claim that the three edge signals plus depth-biased random walk produce paths whose aggregated context measurably improves LLM translation rests on the untested assumption that semantic similarity, adjacency, and keyword overlap suffice to recover the discourse relations (coreference, rhetorical structure, topic shifts) known to matter for DocMT. No section of the manuscript demonstrates that these lexical/local signals recover such relations or that the resulting paths outperform unstructured retrieval of comparable length.
minor comments (1)
  1. The abstract states that experiments were conducted but supplies no information on test sets, metrics (BLEU, COMET, etc.), baseline implementations, or ablation results, preventing assessment of whether the claimed outperformance is supported by the data.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback on the assumptions in our approach. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the three edge signals plus depth-biased random walk produce paths whose aggregated context measurably improves LLM translation rests on the untested assumption that semantic similarity, adjacency, and keyword overlap suffice to recover the discourse relations (coreference, rhetorical structure, topic shifts) known to matter for DocMT. No section of the manuscript demonstrates that these lexical/local signals recover such relations or that the resulting paths outperform unstructured retrieval of comparable length.

    Authors: We agree that the manuscript does not provide a direct demonstration that the chosen edge signals recover specific discourse relations such as coreference or rhetorical structure. Our design is motivated by the intuition that these signals can capture relevant dependencies for context selection in DocMT. The empirical results across multiple domains and LLMs support the effectiveness of the resulting paths. To directly address the concern about outperforming unstructured retrieval, we will include an additional ablation experiment in the revised manuscript that compares the graph-guided paths to length-matched unstructured context selection (e.g., top-k semantic similarity without graph or random selection). This will clarify the contribution of the structured path discovery. revision: yes

Circularity Check

0 steps flagged

No circularity; method is a novel heuristic validated by external empirical comparison.

full rationale

The paper defines a discourse graph from three explicit signals (semantic similarity, adjacency, keyword overlap) and a depth-biased random walk, then feeds sampled paths to an LLM. No equations, fitted parameters, or self-citations are shown that would make any claimed improvement equivalent to the construction itself. The central claim rests on outperformance versus independent baselines across multiple LLMs and domains, which is falsifiable outside the method's own definitions. This is the normal non-circular case for an applied retrieval heuristic.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is limited to the modeling choices explicitly stated there.

axioms (1)
  • domain assumption Paragraph relationships can be adequately captured by a combination of semantic similarity, adjacency, and keyword overlap
    Abstract states this is how the graph nodes are connected.

pith-pipeline@v0.9.1-grok · 5775 in / 1137 out tokens · 24482 ms · 2026-06-28T10:34:15.599742+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 7 canonical work pages · 1 internal anchor

  1. [1]

    In-context examples selection for machine translation

    [Agrawalet al., 2023 ] Sweta Agrawal, Chunting Zhou, Mike Lewis, Luke Zettlemoyer, and Marjan Ghazvinine- jad. In-context examples selection for machine translation. InFindings of the Association for Computational Linguis- tics: ACL 2023, pages 8857–8873,

    2023

  2. [2]

    Graph of thoughts: Solving elaborate problems with large language models

    [Bestaet al., 2024 ] Maciej Besta, Nils Blach, Ales Ku- bicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intel- ligence, volume 38, pages 17682–17690,

    2024

  3. [3]

    Lon- glora: Efficient fine-tuning of long-context large language models.arXiv preprint arXiv:2309.12307,

    [Chenet al., 2023 ] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Lon- glora: Efficient fine-tuning of long-context large language models.arXiv preprint arXiv:2309.12307,

    work page arXiv 2023

  4. [4]

    Efficiently exploring large language mod- els for document-level machine translation with in-context learning.arXiv preprint arXiv:2406.07081,

    [Cuiet al., 2024 ] Menglong Cui, Jiangcun Du, Shaolin Zhu, and Deyi Xiong. Efficiently exploring large language mod- els for document-level machine translation with in-context learning.arXiv preprint arXiv:2406.07081,

    work page arXiv 2024

  5. [5]

    Graft: A graph-based flow-aware agentic framework for document-level machine translation.arXiv preprint arXiv:2507.03311,

    [Duttaet al., 2025 ] Himanshu Dutta, Sunny Manchanda, Prakhar Bapat, Meva Ram Gurjar, and Pushpak Bhat- tacharyya. Graft: A graph-based flow-aware agentic framework for document-level machine translation.arXiv preprint arXiv:2507.03311,

    work page arXiv 2025

  6. [6]

    How to train long-context language models (effectively)

    [Gaoet al., 2025 ] Tianyu Gao, Alexander Wettig, Howard Yen, and Danqi Chen. How to train long-context language models (effectively). InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7376–7399,

    2025

  7. [7]

    Submodular-based in-context example selection for llms- based machine translation

    [Jiet al., 2024 ] Baijun Ji, Xiangyu Duan, Zhenyu Qiu, Tong Zhang, Junhui Li, Hao Yang, and Min Zhang. Submodular-based in-context example selection for llms- based machine translation. InProceedings of the 2024 Joint International Conference on Computational Lin- guistics, Language Resources and Evaluation (LREC- COLING 2024), pages 15398–15409,

    2024

  8. [8]

    BlonDe: An automatic evaluation metric for document-level machine translation

    [Jianget al., 2022 ] Yuchen Eleanor Jiang, Tianyu Liu, Shuming Ma, Dongdong Zhang, Jian Yang, Haoyang Huang, Rico Sennrich, Ryan Cotterell, Mrinmaya Sachan, and Ming Zhou. BlonDe: An automatic evaluation metric for document-level machine translation. InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Li...

    2022

  9. [9]

    [Khandelwalet al., 2021 ] Urvashi Khandelwal, Angela Fan, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis

    Association for Computational Linguis- tics. [Khandelwalet al., 2021 ] Urvashi Khandelwal, Angela Fan, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Nearest neighbor machine translation. InInternational Conference on Learning Representations (ICLR),

    2021

  10. [10]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    [Liuet al., 2023 ] Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language mod- els use long contexts,

    2023

  11. [11]

    Fast nearest neighbor machine translation

    [Menget al., 2022 ] Yuxian Meng, Xiaoya Li, Xiayu Zheng, Fei Wu, Xiaofei Sun, Tianwei Zhang, and Jiwei Li. Fast nearest neighbor machine translation. In Smaranda Mure- san, Preslav Nakov, and Aline Villavicencio, editors,Find- ings of the Association for Computational Linguistics: ACL 2022, pages 555–565, Dublin, Ireland, May

    2022

  12. [12]

    [Phamet al., 2025 ] Viet-Thanh Pham, Minghan Wang, Hao- Han Liao, and Thuy-Trang Vu

    Association for Computational Linguistics. [Phamet al., 2025 ] Viet-Thanh Pham, Minghan Wang, Hao- Han Liao, and Thuy-Trang Vu. Discourse graph guided document translation with large language models.arXiv preprint arXiv:2511.07230,

    work page arXiv 2025

  13. [13]

    [Reiet al., 2022 ] Ricardo Rei, Jos ´e G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and Andr ´e F. T. Martins. COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Philipp Koehn, Lo ¨ıc Bar- rault, Ond ˇrej Bojar, Fethi Bougares, Rajen Chatterjee, Marta R. Costa-juss `a, Christ...

    2022

  14. [14]

    [Thompson and Post, 2020] Brian Thompson and Matt Post

    Association for Computational Linguistics. [Thompson and Post, 2020] Brian Thompson and Matt Post. Paraphrase generation as zero-shot multilingual transla- tion: Disentangling semantic similarity from lexical and syntactic diversity. InProceedings of the Fifth Conference on Machine Translation (Volume 1: Research Papers), On- line, November

    2020

  15. [15]

    [Vernikoset al., 2022 ] Giorgos Vernikos, Brian Thompson, Prashant Mathur, and Marcello Federico

    Association for Computational Lin- guistics. [Vernikoset al., 2022 ] Giorgos Vernikos, Brian Thompson, Prashant Mathur, and Marcello Federico. Embarrassingly easy document-level mt metrics: How to convert any pre- trained metric into a document-level metric. InProceed- ings of the Seventh Conference on Machine Translation, Abu Dhabi, United Arab Emirates,...

    2022

  16. [16]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    Asso- ciation for Computational Linguistics. [Wanget al., 2022 ] Xuezhi Wang, Jason Wei, Dale Schu- urmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171,

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  17. [17]

    Document-level machine translation with large lan- guage models.ArXiv, abs/2304.02210,

    [Wanget al., 2023 ] Longyue Wang, Chenyang Lyu, Tianbo Ji, Zhirui Zhang, Dian Yu, Shuming Shi, and Zhaopeng Tu. Document-level machine translation with large lan- guage models.ArXiv, abs/2304.02210,

    work page arXiv 2023

  18. [18]

    Wong, Fandong Meng, Jie Zhou, and Min Zhang

    [Wanget al., 2025 ] Yutong Wang, Jiali Zeng, Xuebo Liu, Derek F. Wong, Fandong Meng, Jie Zhou, and Min Zhang. DelTA: An online document-level translation agent based on multi-level memory. InThe Thirteenth International Conference on Learning Representations,

    2025

  19. [19]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837,

    [Weiet al., 2022 ] Jason Wei, Xuezhi Wang, Dale Schuur- mans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837,

    2022

  20. [20]

    arXiv preprint arXiv:2401.06468 , year=

    [Wuet al., 2024 ] Minghao Wu, Thuy-Trang Vu, Lizhen Qu, George Foster, and Gholamreza Haffari. Adapting large language models for document-level machine translation. arXiv preprint arXiv:2401.06468,

    work page arXiv 2024

  21. [21]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural informa- tion processing systems, 36:11809–11822,

    [Yaoet al., 2023 ] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural informa- tion processing systems, 36:11809–11822,

    2023

  22. [22]

    In-context example selection via sim- ilarity search improves low-resource machine translation

    [Zebazeet al., 2025 ] Armel Randy Zebaze, Beno ˆıt Sagot, and Rachel Bawden. In-context example selection via sim- ilarity search improves low-resource machine translation. InFindings of the Association for Computational Linguis- tics: NAACL 2025, pages 1222–1252, 2025

    2025