Domain-specific Tunisian Arabic dialect-to-MSA translation in Islamic jurisprudence: a multi-paradigm benchmarking of Seq2Seq models and LoRA-Adapted large language models

The automatic translation of Tunisian Arabic (Darija) into Modern Standard Arabic (MSA) remains a challenging low-resource natural language processing task due to substantial lexical divergence, mixed morphology, code-switching phenomena, and the scarcity of high-quality parallel corpora. These challenges are particularly critical in Islamic jurisprudence ( fiqh ), where translation errors may affect theological and legal meanings. This paper presents a systematic empirical evaluation of parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) for Tunisian Arabic–MSA translation in the religious domain. Three large language models—ALLaM-7B, LLaMA 3-8B, and Mistral-7B—are benchmarked against two sequence-to-sequence baselines (AraT5v2 and AraBART) and zero-shot inference settings using a dedicated corpus of 3,000 Tunisian–MSA sentence pairs extracted from Islamic jurisprudential discourse. Eight experimental configurations are investigated by varying LoRA rank ( r ∈ {16, 32, 64}), regularisation strength, and training duration. Evaluation combines automatic translation metrics (BLEU, chrF++, TER, Token F1, and Exact Match), learned semantic metrics (BLEURT, COMET, and COMETKiwi), statistical significance analysis, and blind human assessment by three native Arabic linguists. The best-performing configuration, ALLaM-7B fine-tuned with LoRA at rank 16, achieves a BLEU score of 80.67, marginally exceeding a fully fine-tuned AraT5v2 model approximately thirty times smaller. However, statistical analyses show that this improvement is not significant, emphasizing the importance of evaluating both statistical and practical gains in low-resource settings. Nevertheless, ALLaM-7B substantially reduces critical translation errors, with independent fiqh expert validation confirming the relevance of these improvements. The results demonstrate that Arabic-centric pre-training is the strongest performance determinant, low LoRA ranks provide better generalisation on small domain-specific corpora, and BLEU-based early stopping effectively mitigates overfitting. Zero-shot LLM performance remains insufficient, confirming the need for domain adaptation for reliable Tunisian Arabic–MSA translation.