A Unique Ranking Framework (URF) with LLM-Based Reranking for English-Hindi Bilingual Corpus Search

Sanjog Arora
JECRC University, Plot No. IS-2036 to 2039, Ramchandrapura Industrial Area, Vidhani, Sitapura Extension, Jaipur 303905, Rajasthan, India
Naveen Hemrajani
Anand International College of Engineering, Near Kanota, Agra Road, Jaipur 303012, Rajasthan, India
Author for correspondence: dean.engineering@jecrcu.edu.in

Abstract

Cross-lingual information retrieval (CLIR) allows users to retrieve relevant information written in a language different from that of the query, and therefore plays a critical role in multilingual knowledge access. Recent advances in dense retrieval and retrieval-augmented generation (RAG) have improved semantic matching; however, current retrieval schemes do not fully align semantic understanding with lexical matching, particularly for morphologically diverse and low-resource language pairs such as Hindi and English. In addition, conventional ranking fusion strategies depend primarily on rank positions and often fail to exploit the underlying semantic and lexical scores.

The present investigation proposes a Unique Ranking Framework (URF) that integrates dense semantic retrieval, BM25-based lexical retrieval and large language model (LLM) based
reranking within a combined RAG architecture. A new bilingual benchmark, BiIR-2026, consisting of 1,860 aligned English-Hindi (En-Hi) document pairs and 742 relevance-judged queries, was constructed from publicly available resources. Experimental evaluation demonstrated that the proposed URF outperformed representative dense, lexical, sparse-neural
and late-interaction retrieval baselines across Precision@10, Recall@10, mean average precision (MAP), mean reciprocal rank (MRR) and normalized discounted cumulative gain at
rank 10 (NDCG@10). Ablation studies further showed that optimised score fusion and selective LLM reranking improved retrieval effectiveness while keeping computational latency feasible for interactive platforms. The proposed framework provides a reproducible and scalable alternative for multilingual retrieval and offers a practical basis for extending RAG to low-resource Indic languages.

Keywords: cross-lingual information retrieval (CLIR), BM25, retrieval-augmented generation (RAG), large language models (LLM), semantic retrieval, hybrid information retrieval, multilingual retrieval, English-Hindi retrieval, vector search, FAISS