An interpretable machine learning framework for classifying human and machine translations across genres.

Neural modeling and large language models (LLMs) has led to a significant improvement in the quality of machine translation (MT) output. While MT increasingly rivals human output, "translationese"-systematic linguistic fingerprints left by the translation process-has long been studied qualitatively, yet its quantitative boundaries remain unclear. We present an interpretable machine-learning framework that classifies Chinese-to-English human, Google-Translate, and ChatGPT outputs across news, novel, and technology genres using a dataset of 450 texts. From 308 candidate linguistic indicators which were normalized for text length, Elastic-Net logistic regression selected 14 robust predictors strictly from the training set to prevent data leakage. The partial least squares discriminant analysis (PLS-DA) model stood out across 10 algorithms and achieved an F1-score of 0.90 and an AUC (0.958 ± 0.022) after bootstrap validation. Global, conditional, and local SHapley Additive exPlanations (SHAP) reveal that normalized discourse-level cohesion (PIN: Present-participial clause density), morphological richness (PRMD: Predictive modal density), and adposition density (d_adp) are the strongest, genre-stable discriminators. Functional analysis suggests that these features serve as proxies for informational density and stancetaking, respectively, allowing the model to distinguish human stylistic sensitivity from machine normalization. These attributions align with theoretical constructs of shining-through and normalization found in Corpus-Based Translation Studies literature. By integrating interpretable modeling, this study addresses the "black box" problem in text classification, offering a potential methodological template for differentiating human from machine translations.