This is a companion discussion topic for the original entry at https://localazy.com/blog/ai-8-llm-arabic-models-tested-to-translate
This is a companion discussion topic for the original entry at https://localazy.com/blog/ai-8-llm-arabic-models-tested-to-translate
One main point, LLM are good to generate text, but not too much in annotating and understanding texts: Arabic text is scarce, but rather because the mix of linguistic, cultural, and technical issues makes it difficult to generate the data needed
I agree with you about the resons above: rich morphology, agglutination and missing diacritics.
Moreover, For grammatical tasks, English is typically annotated with around 60 -70 labels. In contrast, Arabic may require over 1,000 grammatical labels, including approximately 15 POS tags, along with a wide range of morpho-syntactic and orthographic variants resulting from agglutination—excluding any semantic annotations. Theoretically, achieving comparable accuracy in Arabic would require up to 10¹⁴ times more annotated data than in English.
New IEEE Publication
Title: Beyond the Determiner “Al-”: Expanding the Determiner Class in Arabic and Eliminating Lexical Ambiguities through Grammars (25 pages)
LinkedIn
The full text is available, and I recommend downloading the PDF version, which offers a clearer layout than the HTML view.
This work explores the morpho-syntax of determiners such as:
مختلف، معظم، أغلب، كافة، شتى، عدة، بعض، جميع، أحد، إحدى، بضع، بضعة
Our development and evaluation are based on approx. 2,500 / 500 noun phrase chunks extracted from real-world corpora.
We also include a short discussion on LLM hallucinations in the section titled: “A Hybrid Approach: Rule-Based with LLM.”
Feel free to share!