Scholarlink

<p>Arabizi poses persistent challenges for Arabic NLP, particularly for low-resource varieties such as Moroccan Darija. We present an auto-supervised, cycle-consistent transliteration framework that maps Moroccan Darija Arabizi to Arabic script without reliance on large annotated corpora or resource-intensive models. The approach first generates candidate transliterations via Darija-specific phonetic rules, then enforces fidelity through Arabic↔Arabizi round-trip consistency with edit-distance thresholds, subsequently prunes implausible candidates using dialectal lexicons, and finally applies a lightweight multilingual language model (MiniLM) for context-aware re-ranking. Iterative Arabizi→Arabic→Arabizi cycles induce a large pseudo-parallel corpus that further improves model robustness. Evaluated on 210k Moroccan Arabizi words, the method outperforms rule-based, statistical, and neural baselines, achieving 92% word-level accuracy and 87 BLEU, and yields measurable gains for downstream sentiment analysis. The system runs efficiently on CPU, and we release code, lexicons, and the induced corpus to support research in low-resource Arabic dialect processing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Arabizi to Arabic script: auto-supervised transliteration for Moroccan Darija using lightweight phonetic-semantic embeddings

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Abdellah Ait Elouli,
  • Hassan Ouahi,
  • El Mehdi Cherrat,
  • Abdellatif Bekkar

摘要

Arabizi poses persistent challenges for Arabic NLP, particularly for low-resource varieties such as Moroccan Darija. We present an auto-supervised, cycle-consistent transliteration framework that maps Moroccan Darija Arabizi to Arabic script without reliance on large annotated corpora or resource-intensive models. The approach first generates candidate transliterations via Darija-specific phonetic rules, then enforces fidelity through Arabic↔Arabizi round-trip consistency with edit-distance thresholds, subsequently prunes implausible candidates using dialectal lexicons, and finally applies a lightweight multilingual language model (MiniLM) for context-aware re-ranking. Iterative Arabizi→Arabic→Arabizi cycles induce a large pseudo-parallel corpus that further improves model robustness. Evaluated on 210k Moroccan Arabizi words, the method outperforms rule-based, statistical, and neural baselines, achieving 92% word-level accuracy and 87 BLEU, and yields measurable gains for downstream sentiment analysis. The system runs efficiently on CPU, and we release code, lexicons, and the induced corpus to support research in low-resource Arabic dialect processing.