Pith. sign in

REVIEW 5 major objections 5 minor 44 references

GRAFT: A Graph-based Flow-aware Agentic Framework for Document-level Machine Translation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that representing a source document as a directed acyclic graph of discourse segments, with LLM-selected edges and structured memory passed along those edges, gives large and consistent document-level translation gains…

desk verdict A well-built, clearly-documented document-MT pipeline whose headline gains are plausible but not yet distinguishable from noise, because the main benchmark is ten to twelve documents per direction. read the letter →

arxiv 2507.03311 v1 pith:6LWWXLTE submitted 2025-07-04 cs.CL cs.AI

Himanshu Dutta , Sunny Manchanda , Prakhar Bapat , Meva Ram Gurjar , Pushpak Bhattacharyya This is my paper
classification cs.CLcs.AI
keywords document-levelmachinetranslationdiscoursesegmentationdependencygraphLLMagentscontextualmemoryd-BLEUterminologyconsistencyzeropronoun
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

GRAFT claims that document-level machine translation improves when the source document is first rewritten as a directed acyclic graph of discourse segments: an LLM decides where segment boundaries fall, an edge-selection agent decides which earlier segments matter for each later segment, and a memory agent turns each translated segment into structured notes (noun-to-pronoun mappings, entity and phrase translations, discourse connectives, and a one-line summary) that are passed along the graph to downstream segments. On TED test sets spanning eight language directions the full pipeline gains 2.8 d-BLEU on average over strong baselines, and on four Chinese-to-English domains it gains 2.3 d-BLEU over a strong closed-source model. The load-bearing claim is that the chosen edges and the memory they carry, not the LLM backbone, produce the gains: replacing the edge-selection agent with a chain or TF-IDF graph costs 2.0 d-BLEU, and removing the memory mechanism entirely costs 5.7 d-BLEU.

What carries the argument

Four cooperating LLM agents organized as a DAG pipeline. The Discourse Agent decides sentence by sentence whether a sentence continues the current discourse segment; the Edge Agent decides, for each pair of earlier and later segments, whether translation of the later segment needs context from the earlier one, with the immediately preceding segment always included; the Memory Agent extracts a five-component local memory from each translated segment; and the Translation Agent translates each segment using its source text plus the memories of all graph predecessors. The DAG is what carries the argument: it replaces flat or heuristic context with a routable structure whose edges decide which memories a segment sees.

What would settle it

Run GRAFT on the same eight TED directions with the same number of edges but chosen uniformly at random from all earlier segments (keeping the mandatory previous-segment edge), and compare d-BLEU; if random edges close the 2.0 d-BLEU gap over the chain graph, the claim that LLM-selected edges carry the improvement is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the way context is selected and propagated matters as much as the language model behind the translation. GRAFT models a document as a directed acyclic graph whose nodes are LLM-chosen discourse segments and whose edges are LLM-judged dependencies: every segment always depends on its immediate predecessor, and the Edge Agent adds long-range edges when later text needs earlier text for coherent translation. Each translated segment deposits a small structured memory consisting of noun-to-pronoun mappings, entity translations, phrase translations, discourse connectives, and a one-line summary, and each segment is translated using the union of memories from its graph predecessors, with earlier dependencies prioritized. The paper reports that this graph routing, rather than the backbone, drives the gains: on eight language directions the full system beats strong baselines by 2.8 d-BLEU on TED, beats a strong closed-source model by 2.3 d-BLEU on domain-specific Chinese-to-English tests, and loses 2.0 d-BLEU when the graph edges are replaced by a chain or TF-IDF graph.

Load-bearing premise

The Edge Agent's yes/no relevance judgments are assumed to identify the dependencies that actually matter for translation quality; if those judgments are frequently wrong or simply mirror local adjacency, the reported gains could come mostly from the mandatory edge to the previous segment rather than from LLM-selected long-range dependencies.

Editorial extensions

If this is right

  • The reported 2.8 d-BLEU average on TED tst2017 across eight directions, if replicated, makes discourse-graph routing a stronger lever than swapping the LLM backbone for document translation.
  • The 2.3 d-BLEU gain over a strong closed-source model on four Chinese-to-English domains, achieved with a smaller open-weight backbone, suggests domain-specific terminology and zero-pronoun recovery can be improved through structured memory rather than model scale.
  • The 2.0 d-BLEU loss when the DAG edges are replaced by chain or TF-IDF edges implies that the identity of the edges, not just the number of context segments, determines document-level quality.
  • Because removing memory entirely costs 5.7 d-BLEU, the translation agent's later outputs depend on accumulated structured memory in a way that flat windowed context cannot reproduce.
  • The long-novel experiment, translating a whole 16.9K-sentence novel as one document and beating chapter-by-chapter translation by 4.3 d-BLEU, implies the DAG representation scales to very long documents without content omission.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If edge selection is what matters, then distilling the Edge Agent's judgments into a small trained classifier could cut cost while keeping most of the gains; the paper does not test this.
  • The graph idea should transfer to other long-context generation tasks where selective context is key, such as summarization or storytelling, where a document DAG could route relevant earlier content to each new segment.
  • A controlled ablation that matches the number of LLM-selected edges with randomly chosen edges would separate the value of edge choice from the value of simply having more context; this is a direct test the paper does not run.
  • Memory components are not equally valuable: phrase and entity mappings appear to carry more of the gain than the summary or noun-pronoun mappings, so future work could shrink memory to the highest-leverage fields and reduce the reported threefold latency overhead.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. This paper proposes GRAFT, an LLM-agent-based document-level machine translation framework. The source document is segmented into discourse units by a Discourse Agent, a DAG of dependencies is built by an Edge Agent (with mandatory consecutive edges plus LLM-selected non-consecutive edges), a Memory Agent extracts structured memory from each translated segment, and a Translation Agent translates each segment conditioned on memories of its predecessors. The authors report d-BLEU gains on TED tst2017 for eight translation directions (Table 1), on Zh→En domain-specific data across four domains (Table 2), and ablations of the segmentation, edge, and memory components (Tables 5-7), plus human evaluation of discourse phenomena. The headline claims are that GRAFT improves 2.8 d-BLEU over strong baselines on TED and 2.3 d-BLEU on domain-specific translation, and that the DAG structure, not the LLM backbone, drives the gains.

Significance. If correct, the contribution is a structured, interpretable way to select and propagate context in document-level MT, replacing flat or heuristic context aggregation with LLM-mediated segmentation and dependency modeling. The paper is transparent in providing prompt templates, detailed ablations, human evaluation guidelines, and a cost/latency analysis, and it makes its code and data available through an anonymized link. However, the support for the quantitative headline claims is currently insufficient: the main benchmark has only 10-12 documents per direction, the quoted averages are inconsistent across the abstract, contributions, and results sections, and the edge-selection effect is not isolated from the effect of adding edges. The DAG-based representation is a plausible and interesting direction, but the paper needs stronger statistics and internal consistency before the central claim can be accepted.

major comments (5)
  1. [Abstract; Section 4.3; Table 1; Table 2] The headline numbers are not reproducible from the reported tables. The abstract states a 2.8 d-BLEU average improvement over 'strong baselines' on TED, but Table 1 shows GRAFT (Llama-3.1-70B-Instruct) gains of about 0.9 d-BLEU over GPT-4o-mini and about 1.2 d-BLEU over DELTA (Qwen2-72B-Instruct) on average, which matches Section 4.3's 1.0 and 1.1 rather than 2.8. The contributions bullet claims 3.4 d-BLEU over 'similar LLM-based approaches,' which does not appear in any table. In addition, the abstract's domain-specific claim of 2.3 d-BLEU from English to Chinese does not match Table 2, which reports an average 2.0 d-BLEU gain over GPT-4o-mini on Chinese-to-English (Zh→En), not English-to-Chinese. Please define the baseline set for each headline number and reconcile all quoted averages with the tables.
  2. [Table 4; Tables 1, 5, 6, 7] The TED tst2017 benchmark contains only |D|=10-12 documents per translation direction (Table 4), yet all reported d-BLEU scores are corpus-level aggregates with no bootstrap confidence intervals, per-document variance, or paired significance tests. A 1-3 point average difference over 10-12 documents is within sampling noise. Table 6's central comparison (CG/TF-IDF versus Edge Agent) even goes against the claimed trend for Zh→En (24.8 vs 24.6), and Table 1 shows a 17.6-point gain on Fr→En alongside a 2.5-point loss on Ja→En. Without uncertainty quantification, the paper's central causal claim that DAG structure and edge selection drive the improvement cannot be evaluated.
  3. [Eq. (1); Table 6; Appendix G.2] The contribution of the Edge Agent is not isolated. Equation (1) always adds the consecutive edge dj→di for j+1=i, so the Edge Agent only adds non-consecutive edges. Table 6 compares the chain graph (consecutive edges only) with the full Edge Agent (consecutive plus LLM-selected edges); the 2.0 d-BLEU average gain therefore conflates the value of having any additional edges with the value of the LLM's edge selection. The 76.3% human-rated edge accuracy in Appendix G.2 is reported without inter-annotator agreement or error analysis, and no experiments quantify how a wrong edge label degrades translations. Please add a control that matches the number of extra edges (e.g., random edges) and report how edge precision affects translation quality.
  4. [Figure 5; Section 5] The consistency analysis in Section 5 is not supported as reported. Figure 5's second panel is labeled 'Frequency of Consistency Ratios (Simulated Rising Trend),' which suggests the distribution may be simulated rather than measured from GRAFT outputs, yet the text cites it as evidence that 'most paths show a consistency ratio greater than 0.6.' If this panel is an illustration, the surrounding claims must be revised; if it is real data, the label must be corrected. The definition of 'consistency maintained' in CL(P) is also not formalized.
  5. [Section 5; Table 4] The 'ultra-long document' experiment in Section 5 states that the Web Novel subset contains 16.9K source sentences, but Table 4 reports |S|=860 sentences for the Novel row, with 16.9K source words. The claimed scale of the long-range dependency test is therefore unsupported as written. Please correct the sentence/word counts and re-state the experimental scale accordingly.
minor comments (5)
  1. [Section 4.3; Table 1] Section 4.3 refers to 'DELTA (Qwen2.5-72B-Instruct)', while Table 1 lists 'DELTA (Qwen2-72B-Instruct)'; please make the model names consistent.
  2. [Section 4.2; Table 7] The statement that 'the least impactful component, M_np, improves performance by 4.9 d-BLEU for De ⇒ En' is ambiguous: Table 7 shows that removing M_np lowers De⇒En from 31.5 to 26.6 (a 4.9 drop), which makes M_np impactful for that direction even if it is least impactful on average. Please clarify the wording.
  3. [Appendix C.3] The cTT formula appears typeset incorrectly: the denominator 'C2_k TT' is unclear, and the summation indices over terminology words need a precise definition. Please provide a cleaner formulation.
  4. [References] There are several typographical issues in the references, such as 'V oita' for Voita and 'Rethorical structure theory' for Rhetorical Structure Theory. Please proofread the bibliography.
  5. [Figure 5] The right panel of Figure 5 has a truncated y-axis label ('Frequency of Consistency Ratios (Simulated Rising Trend)') and the legend items ('Target Noun Target Pronoun', 'Source Entity Target Entity', 'Source Phrase Target Phrase') are ambiguous; please provide clear captions and legends.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: GRAFT's components are evaluated against external d-BLEU references and ablations; no fitted quantity is renamed as a prediction.

full rationale

The paper's core claim that its DAG representation and graph-based memory propagation improve document-level translation is tested with d-BLEU against human references on IWSLT2017 TED, mZPRT/WMT2022, and Guofeng benchmarks, and against several external baselines including commercial systems and prior LLM-based agents. The Discourse, Edge, Memory, and Translation agents are implemented via few-shot prompting of pretrained LLMs, with no parameters fitted to maximize d-BLEU on the test set; hyperparameters such as temperature are selected heuristically, not by optimizing the reported evaluation metric and then presenting that optimum as a prediction. Ablations compare complete versus incomplete GRAFT configurations empirically, and the edge/segmentation alternatives are also empirical comparisons rather than restatements of the method's definitions. The consistency-ratio analysis is descriptive and does not feed back into training or system selection. No load-bearing self-citation appears; prior-work citations (e.g., DELTA, Hu and Wan) are external baselines. The observed internal inconsistencies in the abstract's headline d-BLEU averages versus Table 1 and the small number of documents per direction are legitimate statistical-reporting concerns, but they are not circularity. Under the review rules, this is a self-contained external evaluation, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

GRAFT introduces no new scientific entity, force, or conserved quantity. Its novel content is a system design: four LLM agents and a DAG memory structure. The ledger items are therefore free parameters and domain assumptions about the reliability of LLM intermediate judgments, the validity of the evaluation metric, and the representativeness of the small test sets. The most important unstated assumption is that the LLM edge detector produces a graph whose errors do not hurt translation, since the paper only reports positive rates (e.g., 76.3% edges accurate) and never analyzes the effect of wrong edges.

free parameters (4)
  • TF-IDF edge threshold tau = not reported numerically
    In Section 4.1, the TF-IDF graph uses an empirically determined threshold tau for edge creation. The value is not given, and the TF-IDF strategy is only a baseline, not part of GRAFT itself, but it is a fitted parameter in the comparison.
  • temperature = 0.1 to 0.3, best at 0.1
    Decoding temperature for Llama-3.1-70B is reported as a tuned hyperparameter (Appendix C.2). The paper says best performance is at 0.1, so it is a small hyperparameter scan rather than a major free parameter.
  • number of few-shot examples = 3
    Three in-context examples are used for each agent prompt; this choice is not ablated.
  • memory size cap = not reported
    The limitations section says performance is sensitive to hyperparameters like memory size, but no memory size cap or its tuning procedure is reported.
assumptions (4)
  • domain assumption An LLM's binary segmentation decision and binary edge relevance decision are reliable enough to improve translation quality.
    The whole pipeline rests on Lseg and Ledge producing useful intermediate structure. The paper gives human ratings (70.4% pronoun resolution, 76.3% edge accuracy) but no agreement measure or error analysis for wrong edges.
  • domain assumption d-BLEU is a valid proxy for document-level translation quality on these test sets.
    The paper uses d-BLEU as the primary metric for all main claims. d-BLEU averages sentence BLEU and may miss document-level improvements such as terminology consistency, so several claims rest on a metric that measures sentence-level overlap.
  • domain assumption The few-shot prompts and examples shown in Appendix I generalize across all eight language directions and domains.
    Prompts are shown only for En to De, and the same templates are used for other directions. The paper does not analyze whether prompt language or example language affects performance in, for example, Ja to En.
  • domain assumption The TED tst2017 set (10 to 12 documents per direction) is representative enough to support claims of consistent gains.
    The main comparison uses very few documents per direction, and no statistical significance tests or confidence intervals are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GRAFT: A Graph-based Flow-aware Agentic Framework for Document-level Machine Translation." pith.science (2026). https://pith.science/paper/6LWWXLTE

@misc{pith2026250703311,
  author       = {Pith},
  title        = {Pith review of: GRAFT: A Graph-based Flow-aware Agentic Framework for Document-level Machine Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6LWWXLTE}},
  note         = {Machine review of arXiv:2507.03311}
}
read the original abstract

Document level Machine Translation (DocMT) approaches often struggle with effectively capturing discourse level phenomena. Existing approaches rely on heuristic rules to segment documents into discourse units, which rarely align with the true discourse structure required for accurate translation. Otherwise, they fail to maintain consistency throughout the document during translation. To address these challenges, we propose Graph Augmented Agentic Framework for Document Level Translation (GRAFT), a novel graph based DocMT system that leverages Large Language Model (LLM) agents for document translation. Our approach integrates segmentation, directed acyclic graph (DAG) based dependency modelling, and discourse aware translation into a cohesive framework. Experiments conducted across eight translation directions and six diverse domains demonstrate that GRAFT achieves significant performance gains over state of the art DocMT systems. Specifically, GRAFT delivers an average improvement of 2.8 d BLEU on the TED test sets from IWSLT2017 over strong baselines and 2.3 d BLEU for domain specific translation from English to Chinese. Moreover, our analyses highlight the consistent ability of GRAFT to address discourse level phenomena, yielding coherent and contextually accurate translations.

Figures

Figures reproduced from arXiv: 2507.03311 by the authors.

Figure 1
Figure 1. An example illustrating a document D segmented into discourse units (di), represented as nodes in a directed acyclic graph. The structure highlights discourse-level phenomena, including sequential connections between discourses, pronoun resolution, discourse connectives, deixis resolution, and the rabbit-hole phe￾nomenon. This representation demonstrates how a document’s internal relationships and cohesive elements … view at source ↗
Figure 2
Figure 2. GRAFT pipeline: illustrating the document-level translation pipeline involving four agents: the Discourse Agent, Edge Agent, Translation Agent, and Memory Agent. The process begins with a document D as input, which is segmented into discourse segments/nodes (di) by the Discourse Agent (Lseg). These segments are then structured into a directed acyclic graph (DAG) with nodes and directed edges, representing dependency… view at source ↗
Figure 3
Figure 3. Analysis of the Discourse Agent with distri [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Assessment of system consistency through [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Task description and prompt: Decision function ( [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Task description and prompt: Edge Agent ( [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Task description and prompt: Memory Agent: [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: Task description and prompt: Memory Agent: [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Prompt: Memory Agent: Mphr i Memory Agent You are a translation memory assistant. Does this German discourse end with a connective like “aber”, “jedoch”, or “deshalb”? If yes, output it; otherwise output . Example 1: Target discourse (DE): Er war müde. Dennoch arbeite…
Figure 11
Figure 11. Figure 11: Task description and prompt: Memory Agent: [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Task description and prompt: Memory Agent: [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Task description and prompt: Translation Agent ( [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 20 canonical work pages

  1. [1]

    Guangsheng Bao, Yue Zhang, Zhiyang Teng, Boxing Chen, and Weihua Luo. 2021. https://doi.org/10.18653/v1/2021.acl-long.267 G -transformer for document-level machine translation . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Lo...

    doi:10.18653/v1/2021.acl-long.267 2021

  2. [2]

    Rachel Bawden, Rico Sennrich, Alexandra Birch, and Barry Haddow. 2018. https://doi.org/10.18653/v1/N18-1118 Evaluating discourse phenomena in neural machine translation . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , pages 1304--1...

    doi:10.18653/v1/n18-1118 2018

  3. [3]

    Mauro Cettolo, Christian Girardi, and Marcello Federico. 2012. https://aclanthology.org/2012.eamt-1.60/ WIT 3: Web inventory of transcribed and translated talks . In Proceedings of the 16th Annual Conference of the European Association for Machine Translation, pages 261--268, Trento, Italy. European Association for Machine Translation

    work page 2012

  4. [4]

    Harrison Chase. 2023. https://python.langchain.com/ Langchain: A framework for developing applications powered by language models . Version as of 2023

    work page 2023

  5. [5]

    Marta R Costa-Juss \`a , James Cross, Onur C elebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672

    arXiv 2022

  6. [6]

    Cade Daniel, Zihang Dai, Yiming Yang, Quanquan Gu, Cho-Jui Hsieh, Denny Zhou, and Cho-Jui Hsieh. 2023. vLLM : A high-throughput and memory-efficient inference and serving engine for large language models. https://github.com/vllm-project/vllm. Accessed: 2025-05-14

    work page 2023

  7. [7]

    Barbara J Grosz and Candace L Sidner. 1986. Attention, intentions, and the structure of discourse. In Computational Models of Discourse, pages 31--51. MIT Press

    work page 1986

  8. [8]

    Chawla, Olaf Wiest, and Xiangliang Zhang

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. https://doi.org/10.24963/ijcai.2024/890 Large language model based multi-agents: A survey of progress and challenges . In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 , pages 8048...

    doi:10.24963/ijcai.2024/890 2024

Show all 44 references
  1. [9]

    Zhiyu Guo and Minh Le Nguyen. 2020. https://doi.org/10.18653/v1/2020.aacl-srw.15 Document-level neural machine translation using BERT as context encoder . In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the ...

    2020 doi

  2. [10]

    Xinyu Hu and Xiaojun Wan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.857 Exploring discourse structure in document-level machine translation . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13889--13902, Singapore. Assoc...

    2023 doi

  3. [11]

    Sebastien Jean, Stanislas Lauly, Orhan Firat, and Kyunghyun Cho. 2017. Does neural machine translation benefit from larger context? arXiv preprint arXiv:1704.05135

    2017 arXiv

  4. [12]

    Linghao Jin, Jacqueline He, Jonathan May, and Xuezhe Ma. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.943 Challenges in context-aware neural machine translation . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15246--15263...

    2023 doi

  5. [13]

    Tom Kocmi, Rachel Bawden, Ond r ej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Thamme Gowda, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Rebecca Knowles, Philipp Koehn, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Michal Nov \'a k, Ma...

    2022

  6. [14]

    Sai Koneru, Miriam Exel, Matthias Huck, and Jan Niehues. 2024. https://doi.org/10.18653/v1/2024.naacl-long.148 Contextual refinement of translations: Large language models for sentence and document-level post-editing . In Proceedings of the 2024 Conference of the North America...

    2024 doi

  7. [15]

    Shaohui Kuang, Deyi Xiong, Weihua Luo, and Guodong Zhou. 2017. https://api.semanticscholar.org/CorpusID:195346279 Cache-based document-level neural machine translation . ArXiv, abs/1711.11221

    2017 arXiv

  8. [16]

    Zongyao Li, Zhiqiang Rao, Hengchao Shang, Jiaxin Guo, Shaojun Li, Daimeng Wei, and Hao Yang. 2025. https://aclanthology.org/2025.coling-main.591/ Enhancing large language models for document-level translation post-editing using monolingual data . In Proceedings of the 31st Int...

    2025

  9. [17]

    Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. https://doi.org/10.1162/tacl_a_00343 Multilingual denoising pre-training for neural machine translation . Transactions of the Association for Computational...

    2020 doi

  10. [18]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2023. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36:46534--46594

    2023

  11. [19]

    William Mann and Sandra Thompson. 1988. https://doi.org/10.1515/text.1.1988.8.3.243 Rethorical structure theory: Toward a functional theory of text organization . Text, 8:243--281

    1988 doi

  12. [21]

    Sameen Maruf, Fahimeh Saleh, and Gholamreza Haffari. 2021 b . https://doi.org/10.1145/3441691 A survey on document-level neural machine translation: Methods and evaluation . ACM Comput. Surv., 54(2)

    2021 doi

  13. [22]

    Lesly Miculicich, Dhananjay Ram, Nikolaos Pappas, and James Henderson. 2018. https://doi.org/10.18653/v1/D18-1325 Document-level neural machine translation with hierarchical attention networks . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Pro...

    2018 doi

  14. [23]

    Hideya Mino, Hitoshi Ito, Isao Goto, Ichiro Yamada, and Takenobu Tokunaga. 2020. https://doi.org/10.18653/v1/2020.coling-main.396 Effective use of target-side context for neural machine translation . In Proceedings of the 28th International Conference on Computational Linguist...

    2020 doi

  15. [24]

    Graham Neubig and Zhiwei He. 2023. Zeno gpt machine translation report. https://hub.zenoml.com/report/1/GPT

    2023

  16. [25]

    Takumi Ohtani, Hidetaka Kamigaito, Masaaki Nagata, and Manabu Okumura. 2019. https://doi.org/10.18653/v1/D19-6505 Context-aware neural machine translation with coreference information . In Proceedings of the Fourth Workshop on Discourse in Machine Translation (DiscoMT 2019), p...

    2019 doi

  17. [26]

    Jianhui Pang, Fanghua Ye, Derek Fai Wong, Dian Yu, Shuming Shi, Zhaopeng Tu, and Longyue Wang. 2025. https://doi.org/10.1162/tacl_a_00730 Salute the classic: Revisiting challenges of machine translation in the age of large language models . Transactions of the Association for ...

    2025 doi

  18. [27]

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22

    2023

  19. [28]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in P ython. Journal of Machine L...

    2011

  20. [29]

    Hongjin Qian, Zheng Liu, Peitian Zhang, Kelong Mao, Defu Lian, Zhicheng Dou, and Tiejun Huang. 2025. https://doi.org/10.1145/3696410.3714805 Memorag: Boosting long context processing with global memory-enhanced retrieval augmentation . In Proceedings of the ACM on Web Conferen...

    2025

  21. [30]

    Yiqi Tong, Jiangbin Zheng, Hongkang Zhu, Yidong Chen, and Xiaodong Shi. 2020. https://doi.org/10.18653/v1/2020.coling-main.388 A document-level neural machine translation model with dynamic caching guided by theme-rheme information . In Proceedings of the 28th International Co...

    2020 doi

  22. [31]

    Zhaopeng Tu, Yang Liu, Shuming Shi, and Tong Zhang. 2018. https://doi.org/10.1162/tacl_a_00029 Learning to remember translation history with a continuous cache . Transactions of the Association for Computational Linguistics, 6:407--420

    2018 doi

  23. [32]

    Elena Voita, Rico Sennrich, and Ivan Titov. 2019. https://doi.org/10.18653/v1/P19-1116 When a good translation is wrong in context: Context-aware machine translation improves on deixis, ellipsis, and lexical cohesion . In Proceedings of the 57th Annual Meeting of the Associati...

    2019 doi

  24. [33]

    Elena Voita, Pavel Serdyukov, Rico Sennrich, and Ivan Titov. 2018. https://doi.org/10.18653/v1/P18-1117 Context-aware neural machine translation learns anaphora resolution . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

    2018 doi

  25. [34]

    Longyue Wang, Siyou Liu, Chenyang Lyu, Wenxiang Jiao, Xing Wang, Jiahao Xu, Zhaopeng Tu, Yan Gu, Weiyu Chen, Minghao Wu, Liting Zhou, Philipp Koehn, Andy Way, and Yulin Yuan. 2024. https://doi.org/10.18653/v1/2024.wmt-1.58 Findings of the WMT 2024 shared task on discourse-leve...

    2024 doi

  26. [35]

    Longyue Wang, Chenyang Lyu, Tianbo Ji, Zhirui Zhang, Dian Yu, Shuming Shi, and Zhaopeng Tu. 2023 a . https://doi.org/10.18653/v1/2023.emnlp-main.1036 Document-level machine translation with large language models . In Proceedings of the 2023 Conference on Empirical Methods in N...

    2023 doi

  27. [36]

    Longyue Wang, Zhaopeng Tu, Yan Gu, Siyou Liu, Dian Yu, Qingsong Ma, Chenyang Lyu, Liting Zhou, Chao-Hong Liu, Yufeng Ma, Weiyu Chen, Yvette Graham, Bonnie Webber, Philipp Koehn, Andy Way, Yulin Yuan, and Shuming Shi. 2023 b . https://doi.org/10.18653/v1/2023.wmt-1.3 Findings o...

    2023 doi

  28. [38]

    Longyue Wang, Zhaopeng Tu, Andy Way, and Qun Liu. 2017 b . https://doi.org/10.18653/v1/D17-1301 Exploiting cross-sentence context for neural machine translation . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2826--2831, Copen...

    2017 doi

  29. [39]

    Wong, Fandong Meng, Jie Zhou, and Min Zhang

    Yutong Wang, Jiali Zeng, Xuebo Liu, Derek F. Wong, Fandong Meng, Jie Zhou, and Min Zhang. 2025. https://openreview.net/forum?id=hoYFLRNbhc Del TA : An online document-level translation agent based on multi-level memory . In The Thirteenth International Conference on Learning R...

    2025

  30. [40]

    Minghao Wu, George Foster, Lizhen Qu, and Gholamreza Haffari. 2023. https://doi.org/10.18653/v1/2023.eacl-main.33 Document flattening: Beyond concatenating context for document-level neural machine translation . In Proceedings of the 17th Conference of the European Chapter of ...

    2023 doi

  31. [41]

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

    2024 arXiv

  32. [42]

    Wong, Hongye Liu, Linfeng Song, Lidia S

    Mingzhou Xu, Longyue Wang, Derek F. Wong, Hongye Liu, Linfeng Song, Lidia S. Chao, Shuming Shi, and Zhaopeng Tu. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.774 G uo F eng: A benchmark for zero pronoun recovery and translation . In Proceedings of the 2022 Conference on E...

    2022 doi

  33. [43]

    Yusen Zhang, Ruoxi Sun, Yanfei Chen, Tomas Pfister, Rui Zhang, and Sercan Arik. 2024. Chain of agents: Large language models collaborating on long-context tasks. Advances in Neural Information Processing Systems, 37:132208--132237

    2024

  34. [44]

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. https://doi.org/10.1609/aaai.v38i17.29936 Expel: Llm agents are experiential learners . In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Confe...

    2024 doi

  35. [45]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  36. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.