Enhancing cross-lingual text classification through linguistic and interpretability-guided attack strategies

Abstract

While adversarial attacks on natural language processing systems have been extensively studied in English, their impact on morphologically complex languages remains poorly understood. We investigate how text classification systems respond to adversarial attacks across Arabic, English, and French — languages chosen for their distinct linguistic properties. Building on the DeepWordBug framework, we develop multilingual attack strategies that combine random perturbations with targeted modifications guided by model interpretability. We also introduce novel attack methods that exploit language-specific features like orthographic variations and syntactic patterns. Testing these approaches on a diverse dataset of news articles (9,030 Arabic, 14,501 English) and movie reviews (200,000 French), we find that interpretability-guided attacks are particularly effective, achieving misclassification rates of 58%–62% across languages. Language-specific perturbations also proved potent, degrading model performance to F1-scores between 0.38 and 0.63. However, incorporating adversarial examples during training markedly improved model robustness, with F1-scores recovering to above 0.82 across all test conditions. Beyond the immediate findings, this work reveals how adversarial vulnerability manifests differently across languages with varying morphological complexity, offering key insights for building more resilient multilingual NLP systems.