Karim Bouzoubaa: Can LLMs Accurately Analyze the Morphology of Classical Arabic Texts?

Описание: Speaker: Karim Bouzoubaa
Title: Can Large Language Models Accurately Analyze the Morphology of Classical Arabic Texts?

Abstract: Today, users increasingly view Large Language Models (LLMs) as reliable and accurate sources of information, a trend reinforced by the growing preference for tools like ChatGPT over traditional search engines. This shift is further supported by the new generations’ inclination toward digital tools over printed books, especially as artificial intelligence continues to gain influence.
From a technical perspective, we have observed that current LLMs struggle to handle questions related to religious topics. Moreover, these models lack a robust understanding of Arabic grammar, which limits their ability to effectively process, interpret, and explain Quranic texts. As a sacred scripture written in Classical Arabic, the Quran offers a rich, open-source corpus comprising over 78,000 words.
Starting from this observation, we defined our main objective: to create a linguistically 100% accurate resource that can be used to train LLMs with high-quality data. Aware of the different levels of linguistic analysis — phonetic, phonological, morphological, syntactic, semantic, and pragmatic — and given that our project focuses on text processing, we chose to begin at the morphological level of Quranic text.
To this end, we adopted a semi-automatic approach to develop our resource, which we named QAMAR. Built using our SAFAR software framework, the resource underwent manual verification over three iterations, conducted over a two-year period. We also compared our results with those produced by other research teams.
In conclusion, our resource clearly outperforms both current LLMs and datasets developed by other scientific teams, while acknowledging the value of their contributions.