Study evaluates hybrid retrieval and LLM-assisted query expansion for Khmer semantic search

Researchers constructed a Khmer semantic search dataset from approximately 17,000 candidate Khmer titles, retaining 3,000 cleaned full-text documents after filtering, normalization, deduplication, and document-length control. The dataset includes 300 manually reviewed user-style Khmer search queries and silver relevance labels with partial human verification.

The study evaluated four retrieval approaches: character n-gram BM25, multilingual dense retrieval, hybrid BM25+dense retrieval, and LLM-assisted query expansion using Qwen2.5 models. BM25 achieved the strongest overall performance, reaching 0.943 recall and 0.876 nDCG. Hybrid BM25+dense retrieval performed comparably, with 0.929 recall and 0.871 nDCG, while dense retrieval alone underperformed relative to both.

LLM-assisted query expansion did not outperform non-expanded retrieval in the experiments. However, the study found that Qwen2.5-3B produced substantially stronger expanded-query results than Qwen2.5-0.5B, suggesting that LLM size and expansion quality are critical factors for low-resource Khmer retrieval.

The analysis also identified risks in direct LLM expansion, including topic drift, generic terms, and noisy reformulations. Simple filtering of expanded queries could inadvertently remove useful semantic cues, underscoring the need for careful post-processing in LLM-assisted retrieval pipelines.

The authors note that their findings highlight both the potential and limitations of LLM-assisted retrieval for Khmer semantic search. They propose that future work focus on stronger human-verified annotations and the development of Khmer-aware retrieval models to address the unique challenges of low-resource languages.

The dataset and documentation are slated for public release at github.com/back-kh/KhmerSemantic-Search.