Analisis Perbandingan Metode Information Retrieval (IR) Pada Implementasi Pencarian Ayat Al-Qur’an

Abstract

The rapid development of artificial intelligence (AI) has driven the transformation of digital services in various fields, including religious services. This study deploys information retrieval methods in the Quranic verse search system by comparing the lexical approach using BM25, the semantic approach through a dense retrieval model embedded intfloat/multilingual-e5-base to generate query vector representations and documents, and connecting the two using the Reciprocal Rank Fusion (RRF) method. Experiments were conducted on a corpus of Quranic verses in JSON format with a top-k search scenario (k=20). System performance was measured using MRR, nDCG, Recall, and MAP metrics. The experimental results show that BM25 has a fairly high relevance coverage with a Recall of 98%. Meanwhile, dense retrieval provides lower performance with an MRR of 60%, nDCG of 64%, and Recall of 81%, thus being below BM25 in most metrics. Merging using RRF yielded the best ranking quality with an MRR of 81%, nDCG of 85%, MAP of 81%, and a consistently high Recall of 97%. These findings demonstrate that the merging/RRF approach provides the best overall performance, while also confirming that the combination of lexical and semantic methods complements each other in improving the relevance and ranking quality of Quranic verse search results.