Machine-learning classification shows that translated Chinese can be separated from originally written Chinese with high accuracy, and that different source languages imprint different 'dialects' on translations into Chinese. But this discriminative success leaves the field's central causal and methodological questions unanswered. This problem, stated as future work by Hu and Kübler (2020) after their multi-source corpus study of translated Chinese, leaves open (i) whether observed marker overuse (e.g. of cohesive markers) can be attributed to the translation process itself rather than to transfer from the source language, using corpora containing the source texts, their translations, and comparable non-translated target-language texts, (ii) what classification-accuracy or graded evidence should count as sufficient for declaring a feature a reliable marker of translationese, given that features significant under population-level t-tests can be at-chance classifiers and vice versa, and (iii) which finer-grained syntactic-complexity measures reconcile information-theoretic indicators (grammar-rule entropy, which suggests higher complexity in translations) with count-based indicators (which suggest simplification), and how they connect to processing or cognitive accounts of translation.

Background

Translations into a language are not statistically indistinguishable from texts originally composed in that language: they form a distinguishable variety commonly called translationese. Two families of explanations compete. Translation universals are hypothesized properties of the translation process itself — most prominently explicitation (the tendency of translations to spell out information that originals leave implicit, classically measured by the frequency of cohesive markers such as conjunctions and pronouns), simplification (translations being lexically and syntactically simpler), and normalization (translations conforming to typical target-language patterns). Source-language interference, by contrast, holds that features of the source text shine through into the translation, so that translations from different source languages carry different imprints — different 'dialects' of translationese into the same target language.

Hu and Kübler (2020) operationalized this debate for Chinese, a typologically distant target under-studied relative to Indo-European languages. Using a corpus of Chinese news translations from seven source languages (German, English, Spanish, French, Japanese, Korean, Russian) and comparable original Chinese news texts from Xinhua, China's state news agency, they ran support-vector-machine classifiers (with 10-fold cross-validation) over linguistically motivated features: a list of 160 cohesive markers compiled by Chen (2006) reached about 96.6% accuracy in distinguishing translations from originals; character and word unigrams reached essentially 100%; the entropy of context-free-grammar (CFG) production rules headed by each phrase type — an information-theoretic complexity measure E(XP)=−∑rp(r)log⁡p(r)E(\mathrm{XP}) = -\sum_{r} p(r) \log p(r) over the rules expanding a phrase type XP — reached about 89.3%. Their conclusions expose three unresolved issues. First, a methodological tension: features such as the frequencies of the ba- construction (a Chinese disposal construction) and the bei- construction (the Chinese passive marker), or total counts of four-character idioms (chengyu) show statistically significant t-test differences between translations and originals yet classify at chance level, while strong classifiers like unigram frequencies carry little theoretical content — no accepted standard says what accuracy (or other evidence) licenses calling a feature a marker of translationese, i.e. evidence for a translation universal rather than a topic or artifact signal. Second, an attribution problem: monolingual comparable-corpus designs (translations vs. originals) cannot separate genuine explicitation by the translation process from source-language influence on the same surface features, because cohesive-marker overuse co-varies with both; resolving this requires tripartite corpora containing the source texts themselves, their translations, and comparable non-translated texts in the target language, controlled for domain and editorial practice — a design long advocated in principle (e.g. Baumgarten, Meyer and Özçetin 2008 on the elusiveness of explicitness) but rarely delivered at scale, and not delivered for the Chinese multi-source setting. Third, an inconsistency in the syntactic-complexity evidence: entropy-based measures indicate higher grammatical complexity in translations for several phrasal categories, while count-based measures (complex noun phrases per clause, mean sentence length) support the simplification hypothesis, and the effect direction varies by phrase type (original Chinese is more complex in noun-phrase structures, translations more complex in other categories) — so the coarse simplification hypothesis needs finer-grained measures linked to cognitive or processing theories of translation. Finally, the authors stress that conclusions from Chinese alone cannot ground claims of universality: cross-typological replication is required. Related results in neighboring settings frame the same tension: Liu and Afzaal (2021) found Chinese–English translations overall less syntactically complex than native English but genre-dependent reversals; Liu et al. (2022) confirmed entropy-based separability of translated from original Chinese; and later work on interpreted Chinese and interpreted English (2025–2026) continues to report mixed simplification/complexification patterns that vary by source language and genre.

Problem Statement

For translated Chinese — and, for universality claims, translations in a typologically representative range of target languages — the question is how to determine which empirically observed translationese features are properties of the translation process itself and which are source-language interference, by constructing and analyzing controlled tripartite corpora that contain (i) source-language originals, (ii) their translations into the target language, and (iii) comparable non-translated texts in the target language, all matched for domain, genre, and editorial practice. On this evidence base, the question decomposes into three components: (1) a feature-by-feature attribution procedure for co-varying features such as cohesive-marker overuse, determining whether increases in explicitness in translations exceed both source-text influence and target-language norms (i.e. whether explicitation as a translation universal is supported once interference is controlled); (2) an explicit, justified evidential standard for the field — what classification-accuracy level, effect size, or graded measure suffices for a feature to count as a reliable marker of translationese and as evidence for a translation universal — resolving the observed tension where features significant under population-level tests (t-tests) classify at chance and high-accuracy classifiers lack theoretical content; and (3) finer-grained syntactic-complexity measures that reconcile information-theoretic indicators (CFG-rule entropy, which shows translations as more grammatically complex for several phrasal categories) with count-based indicators (which support simplification), explain when and why translations become more or less syntactically diverse by phrase type, genre, and source language, and are connected to processing or cognitive theories of translation.

References

  1. Hai Hu and Sandra Kübler. Investigating translated Chinese and its variants using machine learning. Natural Language Engineering 27(3):339–372, 2020. DOI: 10.1017/S1351324920000182. https://doi.org/10.1017/S1351324920000182
  2. Yikang Liu, Wanyang Zhang, Yiming Wang, Jialong Tang, Pei Zhang, Baosong Yang, Fei Huang, Rui Wang, and Hai Hu. Translationese-index: Using Likelihood Ratios for Graded and Generalizable Measurement of Translationese. arXiv:2507.12260, 2025. https://arxiv.org/abs/2507.12260
  3. Daria Pylypenko, Kwabena Amponsah-Kaakyire, Koel Dutta Chowdhury, Josef van Genabith, and Cristina España-Bonet. Comparing Feature-Engineering and Feature-Learning Approaches for Multilingual Translationese Classification. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. DOI: 10.18653/v1/2021.emnlp-main.676. https://doi.org/10.18653/v1/2021.emnlp-main.676
  4. Kanglong Liu and Muhammad Afzaal. Syntactic complexity in translated and non-translated texts: A corpus-based study of simplification. PLOS ONE 16(6):e0253454, 2021. DOI: 10.1371/journal.pone.0253454. https://doi.org/10.1371/journal.pone.0253454
  5. Kanglong Liu, Rongguang Ye, Liu Zhongzhu, and Rongye Ye. Entropy-based discrimination between translated Chinese and original Chinese using data mining techniques. PLOS ONE 17(3):e0265633, 2022. DOI: 10.1371/journal.pone.0265633. https://doi.org/10.1371/journal.pone.0265633
  6. Jia Li and Xianyao Hu. A better or worse communicator? Comparing human and machine translation in source language shining through across registers. Lingua, 312:103834, 2024. DOI: 10.1016/j.lingua.2024.103834. https://doi.org/10.1016/j.lingua.2024.103834
  7. Jia Li and Yunhao Pan. Does language prestige hypothesis hold true across typologically distant languages? A corpus-based study comparing shining-through effect in English-Chinese bidirectional translations. PLOS ONE 21(7):e0353687, 2026. DOI: 10.1371/journal.pone.0353687. https://doi.org/10.1371/journal.pone.0353687
  8. Rui Xie, Wenkang Zhang, Yao Yao, and Andrew K. F. Cheung. Converge or diverge? An entropy-based analysis of source language effects on linguistic simplification in multilingual simultaneous interpreting. Language Sciences, 116:101842, 2026. DOI: 10.1016/j.langsci.2026.101842. https://doi.org/10.1016/j.langsci.2026.101842
  9. Lingxi Fan, Yao Yao, Rui Xie, Chan-in Sio, and Andrew K. F. Cheung. Simplification in interpreting: the text classification of spoken and interpreted Chinese through ensemble learning techniques. Humanities and Social Sciences Communications, 13, 2025. DOI: 10.1057/s41599-025-06382-7. https://doi.org/10.1057/s41599-025-06382-7
  10. Nicole Baumgarten, Bernd Meyer, and Demet Özçetin. Explicitness in translation and interpreting: A critical review and some empirical evidence (of an elusive concept). Across Languages and Cultures 9(2):177–203, 2008. DOI: 10.1556/acr.9.2008.2.2. https://doi.org/10.1556/acr.9.2008.2.2
  11. Parker Riley, Isaac Caswell, Markus Freitag, and David Grangier. Translationese as a Language in 'Multilingual' NMT. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7737-7746, 2020. DOI: 10.18653/v1/2020.acl-main.691. https://doi.org/10.18653/v1/2020.acl-main.691