Hybrid Khmer Semantic Search: An Empirical Comparison of Sparse, Dense, and Fused Retrieval over a Khmer News Corpus

Khmer is a low-resource language whose script lacks explicit word boundaries, which complicates lexical retrieval and leaves the effectiveness of modern dense retrievers largely unverified. This paper reports a controlled, test-collection evaluation of four information-retrieval methods over an offline corpus of 2,054 Khmer news articles and a fixed set of 50 graded-relevance queries spanning three difficulty types. We compare BM25 (sparse), LaBSE (dense), a BM25–LaBSE hybrid fused with Reciprocal Rank Fusion (RRF), and BGE-M3 (a modern multilingual dense model) using Recall@k, MRR@10, MAP@10, and nDCG@10. Contrary to the common assumption that sparse–dense fusion improves retrieval, the BM25–LaBSE hybrid did not outperform BM25 in aggregate (nDCG@10 0.760 vs. 0.756), because the dense component (LaBSE) was the weakest method overall. The strongest method was the standalone dense model BGE-M3 (nDCG@10 0.864), which led on every query type, every aggregate metric, and all three news sources, and whose advantage a Friedman test with Holm-corrected pairwise comparisons confirmed as statistically significant. We conclude that, for formal Khmer news text, the choice of dense model matters more than fusion: a strong off-the-shelf multilingual encoder is the most effective single choice, while fusing a strong sparse signal with a weak dense one yields no measurable benefit. We discuss the principal threat to this conclusion — that the strongest possible hybrid (BM25 + BGE-M3) was not evaluated — and outline it as immediate future work.