Stylistic analysis of translated languages: A perturbation-based XAI deep learning framework

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)
• Machine Learning (ML)

Background:

• Traditional machine learning and deep learning (DL) models are used for text classification in NLP.
• DL models excel at representation learning but often function as black boxes, hindering interpretability.
• Distinguishing translated from non-translated text reveals linguistic style differences.

Purpose of the Study:

• To develop and demonstrate an explainable AI (XAI) deep learning (DL) framework for text classification.
• To interpret the decisions of a DL model used to classify translated and non-translated texts.
• To uncover stylistic differences between translated and non-translated texts beyond superficial features.

Main Methods:

• A variational autoencoder (VAE) was trained using BERT embeddings from translated and non-translated texts.
• A stacked ensemble of three classifiers was used to classify the VAE's latent representations.
• A perturbation-based XAI method was implemented to interpret the DL model's classification decisions.

Main Results:

• The VAE-based model achieved accuracy scores above 0.8 in distinguishing between translated and non-translated texts.
• The XAI analysis successfully interpreted the DL model's decisions.
• Stylistic differences between the text types were identified, extending beyond lexical and syntactic features.

Conclusions:

• The developed XAI DL framework is effective for text classification and model interpretability.
• This framework can uncover nuanced stylistic differences in text data.
• The study demonstrates the potential of XAI DL for various NLP style analysis tasks.