Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Decoding Machine Translationese in English-Chinese News: LLMs vs. NMTs

T0 review · 4 major / 6 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 Machine translation leaves 99%-detectable fingerprints in Chinese news.

desk verdict The LLM-vs-NMT comparison on matched source texts is the credible core; the near-99% OCN-vs-MT claim is undermined by unmatched corpora and in-sample evaluation. read the letter →

arxiv 2506.22050 v1 pith:JXB7ILSL submitted 2025-06-27 cs.CL

Delu Kong , Lieve Macken This is my paper
classification cs.CL
keywords machinetranslationeseEnglish-Chinesetranslationlargelanguagemodelsneuralstylometrydetectionlexicaldiversitynews
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

The paper tries to establish that machine translationese—the systematic stylistic fingerprints machine translation leaves on a text—is present in English-to-Chinese news translation from both neural machine translation systems and large language models, even when the LLM is prompted to write natural Chinese. It shows that a document-level classifier can separate original Chinese news from machine-translated Chinese with about 99% accuracy, meaning machine output is still stylistically distinct from human writing. It also finds that LLMs and NMTs are distinguishable from each other at about 70% accuracy, with LLMs producing more lexically diverse text and NMTs using more brackets. This matters because translationese can affect readability, learning materials, and perceptions of news credibility, and near-perfect detectability challenges claims that LLM translation has reached human parity.

What carries the argument

The argument rides on a five-layer quantitative feature set of 236 features: lexical statistics, part-of-speech and dependency-tag ratios, readability and concreteness scores, translatability markers such as completeness and foreignness, and part-of-speech n-grams. Each feature is expressed as a ratio or weighted measure so text length does not dominate. A chi-square ranking selects the top 30 features per comparison, and those feed five averaged classifiers (Naive Bayes, logistic regression, SVM, decision tree, random forest) plus k-means clustering evaluated with the adjusted Rand index. The features doing the decisive work are sentence length in characters, the ratio of adversative conjunctions, MTLD as the lexical-diversity metric, and the bracket-to-punctuation ratio.

What would settle it

Add professionally human-translated versions of the same 200 English articles to the classification task: if human translations classify as 'original Chinese' as often as machine translations do, or if original Chinese from other outlets is also about 99% separable from the machine group, the claimed translationese signal would instead be a newsroom-style artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that machine translationese is present and detectable in English-to-Chinese news for both NMT engines and LLMs. Using a 236-feature stylometric set and five averaged classifiers, original Chinese news was classified against NMT output at 98.84% accuracy and against LLM output at 98.69% accuracy; combining both machine types into one group yields 98.92% accuracy. The features that most separate original from translated Chinese are sentence length (fewer characters per sentence in machine output, with median about 50 in original versus under 40 in MT) and a higher ratio of adversative conjunctions such as 但是 and 然而. LLMs versus NMTs are separable at 69.38% accuracy, with LLM output showing higher lexical diversity as measured by MTLD and NMT output using more brackets, a habit the authors link to NMT engines appending English proper nouns in parentheses.

Load-bearing premise

The result rests on the assumption that the original Chinese corpus—2,000 texts from People's Daily and Xinhua—is stylistically comparable to the machine-translated Chinese that began as 200 English articles from The Economist and The Guardian, so that consistent differences reflect translation rather than newsroom or genre style.

Editorial extensions

If this is right

  • Machine-generated Chinese news text can be flagged automatically at near-perfect accuracy, which is directly usable in content-integrity and editorial workflows.
  • The comparable OCN-vs-NMT and OCN-vs-LLM accuracies (98.84% versus 98.69%) indicate that prompt-engineered LLM translation has not reduced machine translationese below NMT levels in this language pair.
  • LLM and NMT outputs are distinct enough (about 70% accuracy) to support engine-family attribution in stylometric analysis, even though the two families overlap in clustering.
  • Adversative-conjunction overuse and sentence-shortening are robust markers of English-to-Chinese machine translationese, aligning with earlier findings in other language pairs and supporting a cross-lingual account of translationese.
  • Translation-specific LLMs are not uniformly more human-like: they show lower lexical diversity and more causal conjunctions than generic LLMs, so model design choices trade off different stylometric qualities.

Reading between the lines

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

  • Beyond the paper, the near-perfect OCN-vs-MT separation may partly reflect newsroom style: the 2,000-text original Chinese corpus from People's Daily and Xinhua has a distinctive official register, and each machine sub-corpus is only 200 texts, so a classifier could be learning outlet identity rather than translation status.
  • Beyond the paper, the bracket-usage signal is a plausible diagnostic for NMT-specific proper-noun handling, and testing on corpora heavy in foreign names versus name-free text would confirm whether appending source-language names in parentheses is the underlying mechanism.
  • Beyond the paper, the same five-layer feature pipeline should transfer to other distant language pairs such as English-to-Japanese or English-to-Arabic news, and a cross-lingual replication would show whether sentence-shortening and adversative-conjunction inflation are general translationese universals or specific to English-to-Chinese.
  • Beyond the paper, the absence of a Chinese-versus-foreign LLM difference suggests that translation style is governed more by architecture and training objective than by the LLM developer's language background, which is consistent with the paper's data but is not a claim the paper itself makes.
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

4 major / 6 minor

Summary. The paper investigates machine translationese (MTese) in English-to-Chinese news translation by building a corpus of original Chinese news (OCN) from People's Daily and Xinhua, and machine translations of 200 English news articles from The Economist and The Guardian generated by five NMT engines and six LLMs. Using 236 hand-crafted linguistic features, chi-square feature selection (top-30), five classifiers, and k-means clustering, the paper reports near-perfect classification accuracy for OCN versus MT outputs (98.69–98.92%), about 70% for LLMs versus NMTs, and lower intra-group accuracy. It concludes that MTese is present in both NMT and LLM output, that LLMs show greater lexical diversity and fewer brackets than NMTs, and that translation-specific LLMs differ from generic LLMs in lexical diversity and causal conjunction use.

Significance. If the conclusions were warranted, the paper would be a valuable contribution to translationese research for a distant language pair and one of the first systematic comparisons of LLM and NMT translationese in news. Strengths include a relatively large corpus, a comprehensive five-layer feature set, multiple commercial engines, and a public supplementary repository. However, the central claim that "MTese is present" is not supported by the current design because the original and translated texts come from different outlets, source languages, and topics with no human-translation baseline, and because the reported accuracies are computed in-sample without a held-out validation protocol.

major comments (4)
  1. [Section 3.1 / Table 1] The corpus design confounds translation status with source-language newsroom conventions. OCN consists of 2,000 texts from People's Daily and Xinhua, while every MT sub-corpus is a translation of 200 English texts from The Economist and The Guardian. There is no story alignment, topic matching, or outlet matching, and the Limitations section explicitly concedes that the authors did not conduct sub-genre topic control. Under this design, the near-perfect OCN-versus-MT accuracies in Table 2 (e.g., 98.92% for All Features) may separate state-media Chinese style, outlet-specific conventions, and topic distributions rather than machine translationese. The claim in Section 5.1 that MTese is present is therefore unsupported as stated.
  2. [Section 3.3.2 / Table 2] The classification evaluation is in-sample. No train/test split or cross-validation is described, and the C/T counts in Table 2 equal the full corpus sizes (e.g., 4106/4152 for OCN-MTs, All Features), confirming that the reported accuracies are training-set accuracies. Moreover, chi-square feature selection in Section 3.3.1 is performed on the full dataset before classification, so the reported performance partly measures the fit of the selected features to the same data. The authors should report an evaluation protocol with a held-out set or repeated cross-validation, including the variance across folds and across the five classifiers.
  3. [Section 5.1 / Figure 3] The feature-level comparisons are equally vulnerable to the corpus confound. The observation that MT texts have shorter sentences and more adversative conjunctions than OCN may reflect differences between state-media Chinese news style and translations of Western news style, rather than an effect of the translation process. The acknowledged absence of a human-translation baseline means the authors cannot distinguish machine translationese from translationese in general or from source-text style transfer. The authors should either include a human-translation control or restrict their conclusions to 'differences between original Chinese state-media news and MT output' rather than claiming evidence for MTese per se.
  4. [Section 4.2 / Figure 2] The clustering result is presented as supporting the classification findings, but the ARI of 0.64 indicates only moderate agreement, not the 'clear separation' suggested in the text. The caption mentions 'top-47 shared features' while Section 3.3.3 specifies k=30; the discrepancy is unresolved. More importantly, the same full-corpus feature selection underlies the clustering, so the cluster separation inherits the same optimism and confound as the classification results.
minor comments (6)
  1. [Section 3.3.1] The choice of k=30 for the number of selected features is not justified or varied; a sensitivity analysis (e.g., k=10, 20, 50) would help establish that the conclusions are not an artifact of this specific threshold.
  2. [Section 5.1] The statistic reported for the Kruskal-Wallis test is called an 'F score'; Kruskal-Wallis produces an H statistic, not an F statistic. Please correct the terminology throughout Section 5.
  3. [Figure 3] The phrase 'standardized deviation' appears where 'standard deviation' is intended; this should be corrected for clarity.
  4. [Section 4.1 / Figure 1] The heatmap reports pairwise accuracies averaged over five classifiers, but the spread across classifiers is not reported; given the large range in the heatmap (0.57 to 0.90), the authors should provide per-classifier results or standard deviations so readers can gauge robustness.
  5. [Abstract and Section 5.2] The abstract's phrase 'nearly perfectly distinguishable' is overstated given the moderate ARI of 0.64 and the 70% LLM-NMT accuracy; the wording should be tempered to match the quantitative evidence.
  6. [References] Several references are incomplete or informal (e.g., Chen et al. (2024) is listed as 'arXiv preprint.' with no identifier; Stafanovičs et al. has a formatting issue in the author name). Please check the reference list for consistency.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline OCN-vs-MT separability is an in-sample fit: chi-square feature selection and classifier training are performed on the full labeled corpus with no train/test split, so the near-perfect accuracies partly measure the fitting procedure rather than an independent MTese signal.

  1. fitted input called prediction [Section 3.3.1-3.3.2 and Table 2]
    "a chi-square (χ2) ranking-based feature selection method is employed in both classification and clustering experiments. Features are ranked based on χ2 values, and the top-k features are selected, where k = 30."

    The χ2 ranking uses the class labels of all texts, and the five classifiers are then trained and evaluated on the same full corpus: Table 2 reports C/T = 4106/4152 for OCN-MTs, meaning every training item is also counted as a test item. No train/test split or cross-validation is described anywhere in Section 3.3.2. The reported 98.92% 'all features' accuracy therefore measures how well the fitted features and classifiers separate the already-seen labeled corpus. Presenting this as evidence that original and MT texts are 'nearly perfectly distinguishable' turns an in-sample fit into an independent discriminative result, so the central claim is partly forced by the evaluation loop rather than by out-of-sample evidence.

  2. fitted input called prediction [Section 3.3.3 and Figure 2]
    "The number of clusters (k) is set to 3, and the Euclidean distance is used to measure the similarity between data points. The top-k significant features selected in prior analysis, are utilized as the feature set for clustering."

    The clustering uses the same top-30 features that were selected by supervised χ2 ranking against the labels OCN/LLMs/NMTs, and the resulting clusters are then scored against those same labels with ARI (0.64). The feature-selection step already imports label information into the clustering input, so the 'clear separation' of OCN from MTs in Figure 2 is partly a consequence of selecting features that separate the labeled groups, not an independent unsupervised confirmation of the classification result.

full rationale

The paper contains no mathematical derivation chain, so the definitional-equation, imported-uniqueness, and ansatz-via-citation patterns do not apply. No load-bearing self-citation chain is present. The main circularity is the validation loop in Sections 3.3.1-3.3.3: the χ2 feature selector is run on the full labeled corpus, and the classifiers (and the k-means features) are evaluated on the same corpus, with C/T counts in Table 2 equal to the full sample sizes. The reported near-perfect OCN-vs-MT accuracies are therefore in-sample fit statistics, not out-of-sample predictions, so the strongest claim 'Original Chinese texts are nearly perfectly distinguishable' partly reduces to the fitting procedure. This warrants a score of 6 under the fitted-input-called-prediction pattern. Separately, the corpus design is a serious construct-validity problem: OCN comes from People's Daily and Xinhua, while the MT outputs translate Economist and Guardian articles, with no topic matching and no human-translation baseline. That is a confound rather than a circular reduction, and the paper itself concedes 'nor conduct sub-genre topic control on the news texts selected' and 'we have not included human translations (HTs) in this study.' Those concessions should be weighed as correctness risks, not as additional circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No invented entities. Two hand-chosen hyperparameters (k=30 features, k=3 clusters) and four domain assumptions carry the analysis; the most consequential is the comparability of the OCN and MT corpora.

free parameters (2)
  • feature_selection_top_k = 30
    Section 3.3.1 fixes k=30 for the chi-square-ranked feature set in classification and clustering, without sensitivity analysis or a validation-based selection. The reported accuracies depend on this choice.
  • cluster_count_k = 3
    Section 3.3.3 sets k=3 clusters to match the OCN/LLM/NMT grouping before running k-means; the ARI of 0.64 is conditional on this pre-specified k.
assumptions (5)
  • domain assumption Original Chinese news (People's Daily, Xinhua) and machine-translated Chinese from English news (Economist, Guardian) are comparable enough for attribution of differences to translation status.
    Section 3.1 and Table 1: OCN and MT come from different outlets, topics, and source languages, with no story-level alignment; the conclusion in Section 5.1 that differences are 'MTese' depends on this comparability.
  • domain assumption The LTP toolkit's segmentation, POS tagging, and dependency parsing errors are unbiased across original and translated Chinese, so feature ratios are comparable.
    Section 3.1 states LTP Base2 models are used for segmentation, tagging, and parsing; any systematic error difference between OCN and MT would distort the 236 features.
  • domain assumption Commercial API outputs and the 178-200 texts per engine are representative of each engine's behavior and unaffected by content-filter refusals.
    Table 1 shows several LLMs returned fewer than 200 texts due to 'unsafe content' warnings (footnote 9); if refused texts differ stylistically, sample bias enters the comparisons.
  • domain assumption Averaging results from five off-the-shelf classifiers with default hyperparameters gives a meaningful accuracy estimate.
    Section 3.3.2 uses NB, LR, SVM (linear), DT, and RF with defaults; no tuning or variance estimation is reported.
  • ad hoc to paper Top-k features selected by chi-square on the full dataset are valid inputs for classification without a held-out selection step.
    Section 3.3.1 and 3.3.2: feature selection is performed on the complete corpus before classification; this is not justified and inflates accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoding Machine Translationese in English-Chinese News: LLMs vs. NMTs." pith.science (2026). https://pith.science/paper/JXB7ILSL

@misc{pith2026250622050,
  author       = {Pith},
  title        = {Pith review of: Decoding Machine Translationese in English-Chinese News: LLMs vs. NMTs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JXB7ILSL}},
  note         = {Machine review of arXiv:2506.22050}
}
read the original abstract

This study explores Machine Translationese (MTese) -- the linguistic peculiarities of machine translation outputs -- focusing on the under-researched English-to-Chinese language pair in news texts. We construct a large dataset consisting of 4 sub-corpora and employ a comprehensive five-layer feature set. Then, a chi-square ranking algorithm is applied for feature selection in both classification and clustering tasks. Our findings confirm the presence of MTese in both Neural Machine Translation systems (NMTs) and Large Language Models (LLMs). Original Chinese texts are nearly perfectly distinguishable from both LLM and NMT outputs. Notable linguistic patterns in MT outputs are shorter sentence lengths and increased use of adversative conjunctions. Comparing LLMs and NMTs, we achieve approximately 70% classification accuracy, with LLMs exhibiting greater lexical diversity and NMTs using more brackets. Additionally, translation-specific LLMs show lower lexical diversity but higher usage of causal conjunctions compared to generic LLMs. Lastly, we find no significant differences between LLMs developed by Chinese firms and their foreign counterparts.

Figures

Figures reproduced from arXiv: 2506.22050 by the authors.

Figure 1
Figure 1. Pair-wise comparison of different MT engines [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. K-means clustering using the top-47 shared features, obtained after deduplication of the top-30 salient [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Linguistic differences between OCN and MTs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: reveals that NMT systems use brackets more frequently (average ratio around 0.04) compared to LLMs (around 0.02), as shown by the downward trend in the chart [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Linguistic differences among LTO and other LLMs. The left panel compares MTLD, while the right panel [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 13 canonical work pages

  1. [1]

    Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, and Amanda Askell

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, and Amanda Askell. 2020. https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html Language models are few-shot learners . Advances in neural information processing systems, 33:1877--1901

    work page 2020

  2. [2]

    Sheila Castilho and Natália Resende. 2022. https://doi.org/10.3390/info13020066 Post-editese in literary translations . Information, 13(2):66

    doi:10.3390/info13020066 2022

  3. [3]

    Sheila Castilho, Natália Resende, and Ruslan Mitkov. 2019. https://aclanthology.org/W19-8703/ What influences the features of post-editese? A preliminary study . In Proceedings of the Human - Informed Translation and Interpreting Technology Workshop ( HiT - IT 2019) , pages 19--27

    work page 2019

  4. [4]

    Wanxiang Che, Yunlong Feng, Libo Qin, and Ting Liu. 2021. https://doi.org/10.18653/v1/2021.emnlp-demo.6 N- LTP : an open-source neural language technology platform for Chinese . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing : System Demonstrations , pages 42--49, Online and Punta Cana, Dominican Republic. Associ...

    doi:10.18653/v1/2021.emnlp-demo.6 2021

  5. [5]

    Hao Chen, Abdul Waheed, Xiang Li, Yidong Wang, Jindong Wang, Bhiksha Raj, and Marah I. Abdin. 2024. https://doi.org/10.48550/arXiv.2410.15226 On the diversity of synthetic data and its impact on training large language models . arXiv preprint

    arXiv doi:10.48550/arxiv.2410.15226 2024

  6. [6]

    Petru Cristea and Sergiu Nisioi. 2024. https://aclanthology.org/2024.bea-1.55/ Archaeology at MLSP 2024: machine translation for lexical complexity prediction and lexical simplification . In Proceedings of the 19th Workshop on Innovative Use of NLP for Building Educational Applications ( BEA 2024) , pages 610--617, Mexico City, Mexico. Association for Com...

    work page 2024

  7. [7]

    Joke Daems, Orphée De Clercq, and Lieve Macken. 2017. https://doi.org/10.52034/lanstts.v16i0.434 Translationese and post-editese: how comparable is comparable quality? Linguistica Antverpiensia, New Series – Themes in Translation Studies, 16:89--103

    doi:10.52034/lanstts.v16i0.434 2017

  8. [8]

    Orphée De Clercq, Gert De Sutter, Rudy Loock, Bert Cappelle, and Koen Plevoets. 2021. https://hal.science/hal-03406287/ Uncovering machine translationese using corpus analysis techniques to distinguish between original and machine-translated French . Translation Quarterly, (101):21--45

    2021

Show all 38 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  2. [10]

    Maxim Enis and Mark Hopkins. 2024. https://doi.org/10.48550/arXiv.2404.13813 From LLM to NMT : advancing low-resource machine translation with Claude . arXiv preprint

    2024 doi arXiv

  3. [11]

    Martin Gellerstam. 1986. Translationese in Swedish novels translated from English . Translation studies in Scandinavia, 1:88--95

    1986

  4. [12]

    Zhiwei He, Tian Liang, Wenxiang Jiao, Zhuosheng Zhang, Yujiu Yang, Rui Wang, Zhaopeng Tu, Shuming Shi, and Xing Wang. 2024. https://doi.org/10.1162/tacl_a_00642 Exploring human-like translation strategy with large language models . Transactions of the Association for Computati...

    2024 doi

  5. [13]

    Wei Huang and Haitao Liu. 2009. https://kns.cnki.net/KCMS/detail/detail.aspx?dbcode=CJFD&dbname=CJFD2009&filename=JSGG200929008&v= Application of quantitative characteristics of Chinese genres in text clustering [ In Chinese ] . Computer Engineering and Applications, 45(29):25--27, 33

    2009

  6. [14]

    Yue Jiang and Jiang Niu. 2022. https://doi.org/10.1556/084.2022.00182 A corpus-based search for machine translationese in terms of discourse coherence . Across Languages and Cultures, 23(2):148--166

    2022

  7. [15]

    Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Xing Wang, and Zhaopeng Tu. 2023. https://wxjiao.github.io/downloads/tech_chatgpt_arxiv.pdf Is ChatGPT a good translator? A preliminary study . arXiv preprint arXiv:2301.08745, 1(10)

    2023 arXiv

  8. [16]

    Tom Kocmi, Eleftherios Avramidis, Rachel Bawden, Ondřej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, Barry Haddow, Marzena Karpinska, Philipp Koehn, Benjamin Marie, Christof Monz, Kenton Murray, Masaaki Nagata, Ma...

    2024 doi

  9. [17]

    Lei Lei, Yaoyu Wei, and Kanglong Liu. 2024. AlphaReadabilityChinese : a tool for the measurement of readability in Chinese texts and its applications. Foreign Languages and Their Teaching, 46(1):83--93

    2024

  10. [18]

    Rudy Loock. 2020. https://hal.science/halshs-02913980/ No more rage against the machine: how the corpus-based identification of machine-translationese can lead to student empowerment . Journal of Specialised Translation, (34):150--170

    2020

  11. [19]

    Gerard Lynch and Carl Vogel. 2018. https://doi.org/10.1016/j.csl.2018.05.002 The translator’s visibility: detecting translatorial fingerprints in contemporaneous parallel translations . Computer Speech & Language, 52:79--104

    2018 doi

  12. [20]

    Shushen Manakhimova, Eleftherios Avramidis, Vivien Macketanz, Ekaterina Lapshinova-Koltunski, Sergei Bagdasarov, and Sebastian M \"o ller. 2023. https://doi.org/10.18653/v1/2023.wmt-1.23 Linguistically motivated evaluation of the 2023 state-of-the-art machine translation: Can ...

    2023 doi

  13. [21]

    McCarthy and Scott Jarvis

    Philip M. McCarthy and Scott Jarvis. 2010. https://doi.org/10.3758/BRM.42.2.381 MTLD , vocd- D , and HD - D : a validation study of sophisticated approaches to lexical diversity assessment . Behavior Research Methods, 42(2):381--392

    2010 doi

  14. [22]

    Jiang Niu and Yue Jiang. 2024. https://doi.org/10.1057/s41599-024-02986-7 Does simplification hold true for machine translations? A corpus-based analysis of lexical diversity in text varieties across genres . Humanities and Social Sciences Communications, 11(1):480

    2024 doi

  15. [23]

    Keqin Peng, Liang Ding, Qihuang Zhong, Li Shen, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.373 Towards making the most of C hat GPT for machine translation . In Findings of the Association for Computational Ling...

    2023 doi

  16. [24]

    human parity

    Thierry Poibeau. 2022. https://doi.org/https://hal.science/hal-03738720/ On" human parity" and" super human performance" in machine translation evaluation . In Language Resource and Evaluation Conference , pages 6018--6023

    2022

  17. [25]

    Ming Qian and Chuiqing Kong. 2024. https://aclanthology.org/2024.amta-research.7/ Exploring the advantages and challenges of a concept-guided approach in large language model aided machine translation: integrating generative AI and human-like cognition . In Proceedings of the ...

    2024

  18. [26]

    Mostafizer Rahman, Ariful Islam Shiplu, and Yutaka Watanobe

    Md. Mostafizer Rahman, Ariful Islam Shiplu, and Yutaka Watanobe. 2024. https://doi.org/10.1007/s44196-024-00589-3 CommentClass : a robust ensemble machine learning model for comment classification . International Journal of Computational Intelligence Systems, 17(1):184

    2024 doi

  19. [27]

    Lindsey W. Rowe. 2022. https://doi.org/10.1002/tesq.3143 Google translate and biliterate composing: second‐graders' use of digital translation tools to support bilingual writing . TESOL Quarterly, 56(3):883--906

    2022 doi

  20. [28]

    Artūrs Stafanovičs, Toms Bergmanis, and Mārcis Pinnis. 2020. https://aclanthology.org/2020.wmt-1.73/ Mitigating gender bias in machine translation with target gender annotations . In Proceedings of the Fifth Conference on Machine Translation , pages 629--638, Online. Associati...

    2020

  21. [29]

    Antonio Toral. 2019. https://aclanthology.org/W19-6627/ Post-editese: an exacerbated translationese . In Proceedings of Machine Translation Summit XVII: Research Track, pages 273--281, Dublin, Ireland. European Association for Machine Translation

    2019

  22. [30]

    Antonio Toral, Martijn Wieling, and Andy Way. 2018. https://doi.org/10.3389/fdigh.2018.00009 Post-editing effort of a novel with statistical and neural machine translation . Frontiers in Digital Humanities, 5:9

    2018

  23. [31]

    Eva Vanmassenhove, Dimitar Shterionov, and Matthew Gwilliam. 2021. https://aclanthology.org/2021.eacl-main.188/ Machine translationese: effects of algorithmic bias on linguistic complexity in machine translation . In 16th Conference of the European Chapter of the Associationfo...

    2021

  24. [32]

    Eva Vanmassenhove, Dimitar Shterionov, and Andy Way. 2019. https://aclanthology.org/W19-6622/ Lost in translation: Loss and decay of linguistic richness in machine translation . In Proceedings of Machine Translation Summit XVII: Research Track, pages 222--232, Dublin, Ireland....

    2019

  25. [33]

    Volansky, N

    V. Volansky, N. Ordan, and S. Wintner. 2013. https://doi.org/10.1093/llc/fqt031 On the features of translationese . Digital Scholarship in the Humanities, 30(1):98--118

    2013 doi

  26. [34]

    Longyue Wang, Chenyang Lyu, Tianbo Ji, Zhirui Zhang, Dian Yu, Shuming Shi, and Zhaopeng Tu. 2023. 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 Natu...

    2023 doi

  27. [35]

    Warrens and Hanneke van der Hoef

    Matthijs J. Warrens and Hanneke van der Hoef. 2022. https://doi.org/10.1007/s00357-022-09413-z Understanding the adjusted rand index and other partition comparison indices based on counting object pairs . Journal of Classification, 39(3):487--509

    2022 doi

  28. [36]

    Xu Xu and Jiayin Li. 2020. https://doi.org/10.1371/journal.pone.0232133 Concreteness/abstractness ratings for two-character chinese words in MELD - SCH . PLOS ONE, 15(6):e0232133

    2020 doi

  29. [37]

    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...

  30. [38]

    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.