ColBERT - Static Pruning

This repository is the virtual appendix for the paper Static Pruning for Multi-Representation Dense Retrieval, published in ACM DocEng 2023.

It makes use of PyTerrier and PyTerrier_ColBERT to demonstrate the power of static-pruning methods for reducing ColBERT indices without loss of effectiveness.

Notebooks

This repository contains four notebooks:

• idf-pruning.ipynb - PyTerrier notebook to run the Original, IDF doc-centric and IDF-uniform approaches from Section 5.1, as well as create the indices for Section 5.2.
• baseline-pruning.ipynb - PyTerrier notebook to run the random doc-centric and stopword approaches from Section 5.1.
• pruned-indices.ipynb - PyTerrier notebook to create the runs for the indices in Section 5.2.
• results-analysis.ipynb - PyTerrier notebook to create the figures for Section 5.1.

Index

In order to run the above notebooks, you will require a PyTerrier ColBERT index of the MSMARCO v1 passage ranking corpus. This can be created using the following PyTerrier code:

import pyterrier as pt
from pyterrier_colbert.indexing import ColBERTIndexer
checkpoint="http://www.dcs.gla.ac.uk/~craigm/ecir2021-tutorial/colbert_model_checkpoint.zip"
indexer = ColBERTIndexer(checkpoint, "/path/to/index", "index_name", chunksize=3)
indexer.index(pt.get_dataset("msmarco_passage").get_corpus_iter())
!ln -s /path/to/index/index_name/ivfpq.262144.faiss /path/to/index/index_name/ivfpq.faiss 

Space consumption of the final index is 185GB, as mentioned in the paper.

TREC Covid

We provide a separate notebook providing all indexing and results on the TREC Covid corpus.

Installation

These are necessary:

pip install git+https://github.com/terrierteam/pyterrier_colbert.git
pip install setuptools==59.5.0

Citation

Static Pruning for Multi-Representation Dense Retrieval. Antonio Acquavia, Nicola Tonellotto, Craig Macdonald. In Proceedings of DocEng'2023

@inproceedings{doceng2023,
  title = {Static Pruning for Multi-Representation Dense Retrieval},
  author = {Antonio Acquavia and Craig Macdonald and Nicola Tonellotto},
  booktitle = {Proceedings of ACM DocEng},
  year = {2023},
}

Credits

• Antonio Acquavia, University of Pisa
• Craig Macdonald, University of Glasgow
• Nicola Tonelloto, University of Pisa