pisa-engine/BMP

Faster Learned Sparse Retrieval with Block-Max Pruning. ACM SIGIR 2024.

BMP

Faster Learned Sparse Retrieval with Block-Max Pruning

This repository contains the source code used for the experiments presented in the paper "Faster Learned Sparse Retrieval with Block-Max Pruning" by Antonio Mallia, Torsten Suel and Nicola Tonellotto, published at SIGIR, 2024 - PDF.

Please cite the following paper if you use this code, or a modified version of it:

@inproceedings{BMP,
  author = {Antonio Mallia and Torsten Suel and Nicola Tonellotto},
  title = {Faster Learned Sparse Retrieval with Block-Max Pruning},
  booktitle = {The 47th International ACM SIGIR Conference on Research and Development in Information Retrieval ({SIGIR})},
  publisher = {ACM},
  year = {2024}
}

Usage

Data

The CIFF files and the queries required by BMP to generate an index and perform search operations can be found in the so called CIFF-Hub.

One requirement for BMP to work correctly is that the impact scores of the CIFF files have to be quantized to 8 bits. This is not always done and for this reason is highly recommended to use the CIFF files from the Hub

Index

./target/release/ciff2bmp -b 8 -c ./bp-msmarco-passage-unicoil-quantized.ciff -o bp-msmarco-passage-unicoil-quantized.bmp --compress-range

Search

./target/release/search --index bp-msmarco-passage-unicoil-quantized.bmp --k 1000 --queries dev.pisa > bp-msmarco-passage-unicoil-quantized.dev.trec

Evaluate

trec_eval -M 10 -m recip_rank qrels.msmarco-passage.dev-subset.txt bp-msmarco-passage-unicoil-quantized.dev.trec

Python Bindings

BMP

Install

Form PyPi:

pip intall bmp

From source (in the 'python' directory, i.e cd python):

pip install maturin
maturin build -r
pip install target/wheels/*.whl

Usage

Index

from bmp import ciff2bmp

ciff2bmp(ciff_file="/path/to/ciff", output="/path/to/index", bsize=32, compress_range=False)

Search

from bmp import search, Searcher

# batch operation
results = search(index="/path/to/index", queries="/path/to/queries", k=10, alpha=1.0, beta=1.0)
# -> str (TREC run file)

# query-by-query operation
searcher = Searcher("/path/to/index") # loads index into memory once
searcher.search({'tok1': 5.3, 'tok2': 1.1}, k=10, alpha=1.0, beta=1.0)
# -> Tuple[List[str], List[float]] (doc IDs, scores) for this query