Arabic Authorship Attribution and Style Transfer: New Explorations of Large Language Models on Low-Resource Languages

The benchmark study on Arabic authorship attribution and style transfer conducted by the MBZUAI team has been accepted by LREC 2026. The project has open-sourced its code, models, and datasets, providing an important reference for the application of large language models in low-resource languages and helping to narrow the language gap in AI technology.

Authorship attribution is the task of determining an author's identity based on text, applied in fields such as digital forensics and academic integrity; style transfer is the task of changing the expression style while preserving semantics, suitable for scenarios like content creation and privacy protection. Arabic faces challenges such as linguistic complexity (rich morphology), dialect diversity (differences between Modern Standard Arabic and local dialects), data scarcity (limited annotated corpora), and writing variations (with/without vowel diacritics, etc.). Its research has reference significance for other low-resource languages.

Strategies for adapting LLMs to Arabic include: 1. Using multilingual pre-trained models (e.g., mBERT, XLM-R) for continued pre-training or task-specific fine-tuning; 2. Zero-shot/few-shot learning to address data scarcity issues; 3. Cross-language transfer (translated data, shared representations, adversarial training) to reuse knowledge from high-resource languages.

The MBZUAI team has built a benchmark testing framework for Arabic authorship attribution and style transfer to evaluate the performance of various LLMs; it has open-sourced the complete research code, task-optimized pre-trained models, and dedicated datasets, addressing the long-standing data bottleneck in the field.

The study shows that LLMs still have strong processing capabilities for low-resource languages, bringing hope for narrowing the language digital divide; open-source collaboration and benchmark testing are crucial for promoting the development of the field; the exploration of cross-language methods has reference value for research on other low-resource languages.

Future directions can include exploring Arabic dialect processing, multi-task joint modeling of authorship attribution and style transfer, performance evaluation of larger-scale LLMs, deployment of practical tools, and extending to other low-resource languages to build multilingual benchmarks.

The research results can be applied in scenarios such as digital forensics (tracking the source of anonymous text), academic integrity detection (identifying plagiarism), content creation assistance (adjusting text style), privacy protection (hiding author characteristics), and historical document research (determining anonymous authors).