Cross-Lingual Semantic Search of the Quran Using Multilingual Large Language Models
Cross-Lingual Semantic Search of the Quran Using Multilingual Large Language Models
Maryam Khasheie Varnamkhasti1 Mansooreh Ezhei2 Ahmad Baraani Dastjerdi3
1) PhD student at the Faculty of Computer Engineering, University of Isfahan
2) Assistant Professor at the Faculty of Computer Engineering, University of Isfahan
3) Professor at the Faculty of Computer Engineering, University of Isfahan
IOI Code
:
AAHD-ACZZZH
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
The Quran is a book sent by God to guide millions of Muslims in different languages. Many Muslims use this book daily to search answers for their problems and find peace. The Holy Book Quran is compact in terms of volume, but many supporting books such as tafsirs, hadiths, etc., have been created in various languages to explain its contents. Multilingual semantic search of the Quran across verses, translations, tafsirs, and Islamic texts is one of the major challenges in the field of Information Retrieval.
In this project, cross-lingual semantic retrieval of the Quran and finding verses related to the desired question among verses, translations, and tafsirs for 6,236 Quranic verses in three languages: Arabic, Persian, and English has been made available. For this purpose, the multilingual large language model E5-base has been used, which has been purposefully trained with Quranic data and interpretive texts in the form of pairs of translations and interpretations in different languages. For training the model with limited hardware resources, the LoRA method was used, which enables effective training and preserves semantic accuracy while reducing the number of trainable parameters. Also, by dividing the entire verses into four parts, the model training process was carried out in a staged manner, and ultimately a multilingual Quranic semantic space of three languages was created.
Then, to evaluate the quality of semantic alignment of the languages, embeddings for the translations and interpretations of the three languages were obtained using the trained model, and the cosine similarity between them was calculated. The evaluation metrics results from the constructed shared Quran-centered semantic space are promising. This satisfaction indicates the domain adaptation and the simultaneous use of translations and interpretations in multiple languages. Also, the proposed framework has scalability in adding languages and in training more supporting texts for Quranic verses, which can then be used in real Quran search systems.
Keywords : semantic similarity cross lingual information retrieval semantic search multilingual large language models