Evaluating Cross-Language Information Retrieval Models on Indonesian–Arabic Fiqh Texts: A Case Study

Cross-Language Information Retrieval (CLIR) for highly specialized domains, such as querying classical Arabic jurisprudence (Fiqh) using Indonesian, presents severe vocabulary mismatch and zero-resource training challenges. To resolve this lexical mismatch, we demonstrate that LLM-prompted domain-aware translation successfully captures strict legal terminology where standard machine translation fails. Concurrently, to address the absence of human relevance judgments, we employed the JH-POLO framework to generate synthetic in-domain triplets for fine-tuning a multilingual dense retriever. By synergizing these context-aware sparse signals with the semantic reasoning of the dense bi-encoder via Reciprocal Rank Fusion (RRF), we propose a highly effective hybrid architecture. Empirical evaluations on an expert-curated test collection reveal that while the lexical baseline dominates short queries, this late-fusion pipeline achieves the highest overall accuracy and acts as a robust safety net that consistently maximizes recall for complex, verbose queries.