BM25 outperforms hybrid retrieval for low-resource Khmer semantic search
August 25, 2026
Evaluation of KSE-Web on 3K Khmer documents shows that character n-gram BM25 achieves the strongest performance with 0.943 Recall and 0.876 nDCG. LLM-assisted query expansion using Qwen2.5 was tested against traditional dense and hybrid retrieval methods.
HOW THIS AFFECTS YOU
●
builderFor low-resource languages like Khmer, prioritize BM25 over dense retrieval to ensure better recall and nDCG.
●
researcherThis underscores the continuing gap in multilingual embedding model performance for non-Latin scripts.