The team built a sentence-aligned corpus of Chinese-to-English translations of white papers issued by the State Council Information Office, the government’s official publishing arm. Three parallel sub-corpora were created from the same source texts: human translations completed before 2013 by institutional translators, Google neural machine translation output generated in 2026, and DeepSeek V4Pro translations also produced in 2026 through the model’s official API. Each sub-corpus contained roughly 108,000 to 109,500 words, ensuring comparability. The pre-2013 cutoff for the human translations was deliberate: it guarantees the texts reflect genuine pre-neural institutional practice, untouched by machine post-editing or prompt-driven revision.

To separate stylistic signal from the noise of topic and text length, the researchers applied a rolling stylometry method. Each sub-corpus was treated as a single continuous flow of text, and 5,000-word windows with a 4,500-word step were cut from the human translation series. Because all three versions translate the same Chinese source, the source span underlying each human-translation window was used to mark matching windows in the two machine versions. This produced 24 matched chunk positions per condition, or 72 windows in total, with at most 500 words of overlap between neighboring windows. A diachronic stability check comparing pre-2013 and post-2014 institutional human translations found that none of 136 shared features differed significantly after correction for multiple comparisons, supporting the older corpus as a defensible baseline.

Linguistic features were extracted with the MFTE tagger, which combines part-of-speech tagging with semantic and functional annotation categories, generating 142 measures per window spanning lexical, grammatical, pronominal, modal, and syntactic-complementation levels, all normalized per 1,000 words. Feature selection was run twice to avoid contaminating the evaluation: one pipeline selected 24 features for the exploratory multidimensional analysis using the full sample, while a second, independent pipeline selected 21 features using only the training partition. The researchers then trained five machine-learning classifiers—a linear support vector machine, a random forest, a multilayer perceptron, AdaBoost, and Gaussian Naive Bayes—on a group-aware split that kept every window sharing a chunk identifier entirely on one side of the partition, preventing near-duplicate text from leaking between training and test data.

Naive Bayes and Random Forest emerged as the most consistently reliable classifiers. Naive Bayes achieved a cross-validated accuracy of 0.833 and a test accuracy of 0.800, while Random Forest reached 0.867 on the held-out test set, both far above the 33.3 percent chance baseline for a three-way task. The other classifiers were weaker or more sensitive to the specific data partition: the multilayer perceptron and AdaBoost scored near chance on the single held-out split, though nested cross-validation and 50 repeated group-aware splits showed the perceptron’s performance was better than that single result suggested, with higher variability. A conservative check that excluded training chunks adjacent to test chunks modestly reduced accuracy for most classifiers, indicating a small residual contribution from the 500-word overlap. The overall conclusion was qualified but clear: linguistic form carries genuine, above-chance information about which agent produced a translation.

The exploratory multidimensional analysis, following Biber’s factor-analytic framework, distilled the 24 significant features into five functional dimensions accounting for 59.2 percent of total variance. The standout was Dimension 3, defined by discourse particles, it-cleft constructions, causal conjunctions, demonstratives, and activity verbs. This dimension captures explicit, argumentatively structured cohesion—the connective tissue that links claims, reasons, and concessions into a sustained line of reasoning. It was by far the strongest and most stable group contrast, with a Kendall’s W effect size of 0.814 and significance at p < 0.001. Human translation scored highest, Google NMT next, and DeepSeek V4Pro lowest, with all three pairwise contrasts surviving Holm correction.

The remaining four dimensions were individually weaker. Dimension 1 captured nominal and lexical density—longer, compound-heavy noun phrases—and produced a result that reverses the naive expectation that machines simplify: both machine systems, and DeepSeek in particular, showed greater nominal density than human translation. The researchers attribute this plausibly to “shining through,” the transfer of source-language structure into the translation; Chinese is nominally dense, and machine systems optimizing closer to the source carry more of that density into English, while professional translators re-verbalize dense nominal strings into more idiomatic English clauses. Dimension 2 grouped nominalization, third-person reference, and wh-subordination; Dimension 4 combined lexical diversity with progressive aspect; and Dimension 5 covered elaborative connectives and predicative complexity. On Dimensions 1, 2, and 4, human translation differed significantly from both machine systems, while the two machine systems were statistically indistinguishable from each other.

A non-circularity check addressed a subtle methodological concern: because the exploratory features were selected precisely because they differ across the three conditions, re-testing those features for group differences carries an element of circularity. The researchers re-derived the factor structure from an independently selected, variance-based feature subset and compared the two solutions using Procrustes matching. After matching, all five dimensions were again significant, and the direction of pairwise contrasts replicated in 12 of 15 comparisons. Congruence coefficients were strong for Dimensions 1, 2, and 4, more modest for Dimension 5, and lowest for Dimension 3 at 0.382—yet the discourse-cohesion contrast itself remained significant after matching. The authors therefore treat the full five-factor solution as exploratory, while recommending the discourse-cohesion contrast as the most secure finding to carry forward.

To answer whether DeepSeek’s output is stylistically distinct from conventional neural machine translation, the team computed Euclidean distances between all window pairs in the five-dimensional factor space and assessed uncertainty with a chunk-level bootstrap of 5,000 resamples. The mean distance between NMT and DeepSeek windows was 2.988, compared with 3.251 for human-to-NMT and 3.408 for human-to-DeepSeek. Both human-machine distances exceeded the machine-machine distance with confidence intervals excluding zero, meaning the two machine systems sit reliably closer to each other than either does to human translation. The difference between the two human-machine distances, however, spanned zero, so the data do not support any claim that DeepSeek is detectably more or less human-like than Google NMT in this space.

The authors are careful about interpretation. Because the study included no reader experiment, the suggestion that reduced cohesive marking increases readers’ processing load—and that missing connectives in political texts could shift how confidently readers attribute positions to the issuing government—is presented as a hypothesis grounded in cognitive models of writing, not a demonstrated outcome. Likewise, an account linking large language model behavior to reinforcement learning from human feedback, which optimizes outputs for general conversational helpfulness rather than register-specific institutional norms, is flagged as theoretically motivated but untested here. The study also does not evaluate translation accuracy or fidelity: the human translations are published, post-edited texts while the machine outputs are raw, so the defensible claim is a difference in stylistic profile, not quality. Limitations include the single text type and source, the decade-spanning human baseline, the single DeepSeek model version, and the specific sentence-numbered prompting protocol used. The practical implication the authors draw is targeted: post-editing of machine-translated government text should focus on restoring explicit discourse cohesion—reinserting causal, contrastive, and resultative connectors—rather than applying uniform corrections, and documents whose communicative success depends heavily on argumentative scaffolding may benefit most from human translation or focused post-editing regardless of which machine system produced the draft.

Subject of Research: Stylometric comparison of DeepSeek, Google NMT, and human translations of Chinese government white papers

Article Title: Differentiating DeepSeek AI-generated translations from NMT translations and human translations within Chinese white paper discourse: a computer-based method for analyzing writing style

Article References: Zhang, R., Ji, X., & Zou, J. (2026). Differentiating DeepSeek AI-generated translations from NMT translations and human translations within Chinese white paper discourse: a computer-based method for analyzing writing style. Frontiers in Psychology, 17, Article 1931933. https://doi.org/10.3389/fpsyg.2026.1931933

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1931933

Keywords: machine translation, large language models, DeepSeek, neural machine translation, translationese, computational stylistics, multidimensional analysis, Chinese white papers, discourse cohesion, corpus linguistics, translation studies, stylometry