Arabic Legal Question Classification Using Deep Learning and Adapted Word Embedding

Natural language processing (NLP) techniques have shown promising applications in the legal domain. However, Arabic remains an under-resourced language, particularly for legal text analysis. Despite the potential benefits of applying NLP to Arabic legal texts, such as improved efficiency for legal professionals through automated question classification and answering systems, research in this area remains limited. This paper proposes a novel approach to classify Arabic legal questions relevant to the Moroccan legal domain. A new Arabic legal dataset was curated and used to train domain-specific word embedding models called LegalVec and LegalFastText. Four deep learning architectures were then employed for the legal question classification task, utilizing both our domain-specific embeddings and pre-trained embeddings. The results demonstrate that using our proposed model with our word embeddings LegalVec and legalFastText significantly improves classification performance compared to open-domain pre-trained embeddings. The model that combines gated recurrent unit (GRU) architecture with LegalFastText embeddings was the most efficient achieving 70.95% accuracy. This work represents a step towards developing robust NLP systems tailored to the Arabic legal domain.