SPARe: Supercharged Lexical Retrievers on GPU with Sparse Kernels

Lexical sparse retrievers, rely on efficient searching algorithms that operate over inverted index structures, tailored specifically for CPU. This CPU-centric design poses a challenge when adapting these algorithms for highly parallel accelerators, such as GPUs, thus deterring potential performance gains. To address this, we propose to leverage the recent advances in sparse computations offered by deep learning frameworks to directly implementing sparse retrievals on these accelerators. This paper presents the SPARe (SPArse Retrievers) Python package, which provides a high-level API to deal with sparse retrievers on (single or multi)-accelerators, by leveraging deep learning frameworks at its core. Experimental results show that SPARe, running on an accessible GPU (RTX 2070), can calculate the BM25 scores for close to 9 million MSMARCO documents at a rate of 800 questions per second with our specialized algorithm. Notably, SPARe proves highly effective for denser LSR indexes, significantly surpassing the performance of established systems such as PISA, Pyserini and PyTerrier. SPARe is publicly available at https://github.com/ieeta-pt/SPARe .