The Role of Orthographic Consistency in Multilingual Embedding Models for Text Classification in Arabic-Script Languages

TL;DR

This paper introduces language-specific RoBERTa models for Arabic-script languages, demonstrating that script-focused pre-training improves text classification performance over general multilingual models.

Contribution

The paper presents the AS-RoBERTa family of models, pre-trained on language-specific corpora, to better capture orthographic features and improve classification results in Arabic-script languages.

Findings

AS-RoBERTa outperforms mBERT and XLM-RoBERTa by 2-5% in classification accuracy.

Script-focused pre-training is crucial for capturing language-specific orthographic patterns.

Error analysis reveals the impact of shared script traits and domain content on model performance.

Abstract

In natural language processing, multilingual models like mBERT and XLM-RoBERTa promise broad coverage but often struggle with languages that share a script yet differ in orthographic norms and cultural context. This issue is especially notable in Arabic-script languages such as Kurdish Sorani, Arabic, Persian, and Urdu. We introduce the Arabic Script RoBERTa (AS-RoBERTa) family: four RoBERTa-based models, each pre-trained on a large corpus tailored to its specific language. By focusing pre-training on language-specific script features and statistics, our models capture patterns overlooked by general-purpose models. When fine-tuned on classification tasks, AS-RoBERTa variants outperform mBERT and XLM-RoBERTa by 2 to 5 percentage points. An ablation study confirms that script-focused pre-training is central to these gains. Error analysis using confusion matrices shows how shared script traits and domain-specific content affect performance. Our results highlight the value of script-aware specialization for languages using the Arabic script and support further work on pre-training strategies rooted in script and language specificity.