Faster Learned Sparse Retrieval with Block-Max Pruning

TL;DR

The paper tackles the efficiency gap of learned sparse retrieval, where inverted indexes built from contextualized sparse representations challenge traditional pruning strategies. It introduces Block-Max Pruning (BMP), a two-phase approach that uses block filtering to compute block-level upper bounds and a selective block-evaluation process with tunable early termination ($\alpha$) and optional query-term pruning ($\beta$), operable on a forward/hybrid index. BMP demonstrates substantial speedups (up to ~60x) over prior methods across SPLADE, ESPLADE, and uniCOIL on MS MARCO for safe retrieval, and achieves superior latency-quality tradeoffs for approximate retrieval, often with sub-millisecond responses and minimal RR@10 loss. The work significantly improves the practicality of deploying learned sparse IR at scale and motivates future comparisons with graph-based ANN methods for sparse retrieval.

Abstract

Learned sparse retrieval systems aim to combine the effectiveness of contextualized language models with the scalability of conventional data structures such as inverted indexes. Nevertheless, the indexes generated by these systems exhibit significant deviations from the ones that use traditional retrieval models, leading to a discrepancy in the performance of existing query optimizations that were specifically developed for traditional structures. These disparities arise from structural variations in query and document statistics, including sub-word tokenization, leading to longer queries, smaller vocabularies, and different score distributions within posting lists. This paper introduces Block-Max Pruning (BMP), an innovative dynamic pruning strategy tailored for indexes arising in learned sparse retrieval environments. BMP employs a block filtering mechanism to divide the document space into small, consecutive document ranges, which are then aggregated and sorted on the fly, and fully processed only as necessary, guided by a defined safe early termination criterion or based on approximate retrieval requirements. Through rigorous experimentation, we show that BMP substantially outperforms existing dynamic pruning strategies, offering unparalleled efficiency in safe retrieval contexts and improved tradeoffs between precision and efficiency in approximate retrieval tasks.

Figures (1)

• Figure 1: Effectiveness-efficiency graph for different query processing algorithms. Every line corresponds to an algorithm, every point corresponds to a different configuration.