H3D benchmark pits MinHash vs BGE for fine-grained document deduplication

A team of researchers has released H3D, a comprehensive benchmark for unsupervised text hashing specifically designed for fine-grained document deduplication. The benchmark evaluates seven methods across two categories: non-learning approaches (MinHash, SimHash, Winnowing, FuzzyHash, FlyHash) that generate hash fingerprints via manually designed rules without training, and semantic-sensitive methods built from frozen BGE embeddings with two quantization strategies (BGE-BIHash and BGE-LSHash). All methods were tested on two datasets offering complementary evaluation settings: CSFCube for facet-level scientific-document similarity and RELISH for larger-scale split-level biomedical similarity search.

The results reveal a consistent trade-off between lexical/structural fingerprints and semantic-sensitive representations. Non-learning methods are competitive for near-duplicate matching, while semantic methods better preserve similarity under content rewriting but at significantly higher computational cost. H3D jointly reports ranking quality (MAP, NDCG@20), efficiency, and robustness under controlled text compression. The analysis also identifies when different similarity measures become rank-equivalent for specific hash representations, improving interpretability and reproducibility of method comparisons.