Information Retrieval (IR) systems are crucial in the development of accurate and relevant search engines, especially in domain-specific applications such as Sequran. While fine-tuning is a common optimization approach, this method often requires significant computational resources, diverse datasets, and time-consuming hyperparameter tuning, with the risk of performance degradation. To address these challenges, this research introduces and evaluates a novel architecture that enhances search relevance by integrating lexical and semantic signals, offering a practical alternative to resourceintensive fine-tuning. The proposed architecture involves a two-stage process. The first stage, composite scoring, enhances term frequency (BM25S) with a semantic intent booster. The second stage utilizes a Cross-Encoder (jina-reranker-v2-base-multilingual) to refine the ranking based on contextual relevance. Evaluation is conducted on a domain-specific Islamic query-answer dataset consisting of approximately 20 queries and around 18,000 text entries (over 6,000 verses, each paired with two tafsir types). The proposed method achieves Precision@10 = 0.230 and Recall@10 = 0.564 over the fine-tuned baseline (0.167/0.423), representing absolute gains of 6.3 and 14.1 percentage points, equivalent to relative improvements of 37.7% and 33.3%. Gains are consistent across both metrics, with the larger recall gain indicating that reranking surfaces relevant passages ranked below the cut-off by lexical matching alone. This improvement demonstrates a clear trade-off, as the total execution time increased to approximately 18.5 seconds, which may constrain realtime use. The main implication of this research is the validation of a practical architecture for improving IR systems, offering a viable alternative for domain-specific contexts such as Sequran.
Afiliasi berasal dari OpenAlex; metadata OAI-PMH jurnal tidak memuatnya. Penulis tanpa ID OpenAlex dicocokkan berdasarkan nama, sehingga dua orang bisa tergabung atau satu orang bisa terpisah.