Scholarlink

Arabic text diacritization (tashkil) remains challenging for automated systems. We present a comprehensive comparative evaluation of self-hosted neural networks (CATT), open-source LLMs (Command-R7B), linguistic models (CAMeL Tools), and commercial solutions (Gemini 2.5 Flash Lite, GPT-5-nano, Claude models) on a diverse 1,200-sentence corpus including Classical Islamic text, modern prose, and news. For privacy-sensitive deployments: CATT achieves 3.64% DER with complete data sovereignty as the best self-hosted option. For cost-effective deployments: Gemini 2.5 Flash Lite achieves 2.63% DER at ultra-low cost ($0.10–0.40 per 1M tokens), and GPT-5-nano offers 3.35% DER for budget-constrained applications. For accuracy-critical applications: Claude-Opus-4.5 achieves 2.58% DER as a premium option. Command-R7B-Arabic exhibited severe limitations (25.76% DER, 73% word mismatch rate) due to truncation and hallucination. Our findings reframe the self-hosted-commercial debate: with ultra-affordable commercial solutions available, deployment choices are now primarily about privacy and accuracy priorities rather than cost alone.

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

Evaluating Arabic Diacritization Models: From Self-hosted Neural Networks to Cost-Effective Commercial Solutions

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Khaldoun Senjab,
  • Mohammed Lataifeh,
  • Ashraf Elnagar

摘要

Arabic text diacritization (tashkil) remains challenging for automated systems. We present a comprehensive comparative evaluation of self-hosted neural networks (CATT), open-source LLMs (Command-R7B), linguistic models (CAMeL Tools), and commercial solutions (Gemini 2.5 Flash Lite, GPT-5-nano, Claude models) on a diverse 1,200-sentence corpus including Classical Islamic text, modern prose, and news. For privacy-sensitive deployments: CATT achieves 3.64% DER with complete data sovereignty as the best self-hosted option. For cost-effective deployments: Gemini 2.5 Flash Lite achieves 2.63% DER at ultra-low cost ($0.10–0.40 per 1M tokens), and GPT-5-nano offers 3.35% DER for budget-constrained applications. For accuracy-critical applications: Claude-Opus-4.5 achieves 2.58% DER as a premium option. Command-R7B-Arabic exhibited severe limitations (25.76% DER, 73% word mismatch rate) due to truncation and hallucination. Our findings reframe the self-hosted-commercial debate: with ultra-affordable commercial solutions available, deployment choices are now primarily about privacy and accuracy priorities rather than cost alone.