Stylistic analysis of translated languages: A perturbation-based XAI deep learning framework
Clinical Snapshot
PICO Framework
| P — Population | Not applicable in a clinical sense; the study corpus consists of United Nations meeting transcripts comprising translated and non-translated English texts |
| I — Intervention | A deep learning framework combining a Variational Autoencoder (VAE) with BERT embeddings, a stacked ensemble classifier (majority vote from three classifiers), and a perturbation-based Explainable AI (XAI) method |
| C — Comparator | Traditional machine learning approaches to translated language detection; non-translated UN meeting texts serve as the reference class |
| O — Outcomes | Classification accuracy (>0.8) in distinguishing translated from non-translated text; identification of stylistic features driving model decisions via XAI interpretation |
Bottom Line
This proof-of-concept computational linguistics study presents a three-stage deep learning framework — combining Variational Autoencoders with BERT embeddings, a stacked ensemble classifier, and perturbation-based Explainable AI — to distinguish translated from non-translated text in United Nations meeting transcripts, achieving accuracy above 0.8. The primary contribution is methodological: demonstrating that XAI techniques can render opaque deep learning models interpretable in stylometric tasks, revealing stylistic differences beyond surface lexical and syntactic features. For clinicians and health informaticists, the direct applicability is currently limited. The study is confined to a single highly specialised corpus, reports only accuracy without precision, recall, confidence intervals, or statistical testing, and provides no external validation. Corpus selection bias, source language confounds, and the absence of expert linguistic validation of XAI outputs are notable limitations. The proof-of-concept framing is appropriately modest. Future work validating this framework across diverse text domains — including clinical records, patient-facing health information, and interpreter-mediated consultation transcripts — could yield meaningful applications for translation quality assurance in linguistically diverse healthcare settings. At present, this study represents an early-stage methodological demonstration rather than a practice-changing finding.
Key Findings
P Value: Not reported
Effect Size: Accuracy scores above 0.8 (specific values not reported in abstract)
Primary Outcome: Classification accuracy in distinguishing translated from non-translated UN meeting texts using the VAE-BERT-ensemble framework
Nnt Or Sensitivity: Sensitivity, specificity, precision, recall, and F1-score not reported in abstract; framework performance characterised solely by accuracy threshold
Confidence Interval: Not reported
Clinical Application
The framework requires substantial computational infrastructure (BERT embeddings, VAE training, stacked ensemble) and NLP expertise. Feasibility in routine clinical or health service settings is currently low without significant technical adaptation and resource investment. This study has no direct relevance to Australian PBS, TGA regulatory frameworks, or RACGP clinical guidelines. Indirect relevance may exist for Australian health services serving linguistically diverse communities (CALD populations), where translation quality of clinical communications, patient information materials, and interpreter-mediated consultations is a recognised patient safety concern. The National Accreditation Authority for Translators and Interpreters (NAATI) and the Australian Commission on Safety and Quality in Health Care's work on health literacy in CALD populations represent the closest policy context. Any application to clinical translation quality assurance would require substantial further validation. This study has no direct clinical patient population. Potential downstream relevance exists for health informatics, clinical NLP, and medical translation quality assurance contexts where distinguishing translated from original clinical text may affect information fidelity.
Abstract
Text classification using traditional machine learning techniques has been used in natural language processing (NLP) tasks to distinguish translated from non-translated languages, with high accuracy scores indicating the distinctive style of translated languages. While deep learning (DL) has demonstrated impressive performance in terms of representation learning and capturing nuanced patterns in natural language data, DL models act as black boxes, making their results difficult to interpret. This study addresses this issue by demonstrating an explainable AI (XAI) DL framework in a case study of United Nations (UN) meetings. The framework consists of three stages: i) train a variational autoencoder (VAE) combined with BERT embeddings converted from translated and non-translated texts; ii) utilize the majority vote from three classifiers selected from a stacked ensemble to classify the VAE's latent representations; iii) implement a perturbation-based XAI method to interpret the DL model's decisions. The results indicate that the VAE-based model effectively distinguishes the two text types, with accuracy scores above 0.8. The XAI analysis reveals that interpreting the VAE-based model's decision uncovers stylistic differences between the two text types beyond superficial lexical and syntactic features. This proof-of-concept study demonstrates the potential of the XAI DL framework in other NLP studies that aim to analyze style.
References
- 1.Huang, D. F., & Tay, D. (2026). Stylistic analysis of translated languages: A perturbation-based XAI deep learning framework. PLoS ONE. https://doi.org/10.1371/journal.pone.0352889
Related Research
Journal of medical Internet research
Advancing Human-Centered AI in Clinical Decision Support: Sociocognitive Human-in-the-Loop Study in HIV Care
2 Aug 2026
Neurological research
Federated deep learning model for epilepsy seizure detection using electroencephalogram (EEG) signal
2 Aug 2026
Neural networks : the official journal of the International Neural Network Society
MFBD: Model-free backdoor defense based on vision-language pre-trained models
2 Aug 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service