pyterrier-colbert & ColBERT-PRF

Advanced PyTerrier bindings for ColBERT, including for dense indexing and retrieval. This also includes the implementations of ColBERT PRF, approximate ANN scoring and query embedding pruning.

Usage

Given an existing ColBERT checkpoint, an end-to-end ColBERT dense retrieval index can be created as follows:

from pyterrier_colbert.indexing import ColBERTIndexer
indexer = ColBERTIndexer("/path/to/checkpoint.dnn", "/path/to/index", "index_name")
indexer.index(dataset.get_corpus_iter())

An end-to-end ColBERT dense retrieval pipeline can be formulated as follows:

from pyterrier_colbert.ranking import ColBERTFactory
pytcolbert = ColBERTFactory("/path/to/checkpoint.dnn", "/path/to/index", "index_name")
dense_e2e = pytcolbert.end_to_end()

A ColBERT re-ranker of BM25 can be formulated as follows (you will need to have an index with text saved - the Terrier data repostiory conviniently provides such an index):

bm25 = pt.terrier.Retriever.from_dataset('msmarco_passage', 'terrier_stemmed_text', wmodel='BM25', metadata=['docno', 'text'])
sparse_colbert = bm25 >> pytcolbert.text_scorer()

Thereafter it is possible to conduct a side-by-side comparison of effectiveness:

pt.Experiment(
    [bm25, sparse_colbert, dense_e2e],
    dataset.get_topics(),
    dataset.get_qrels(),
    eval_metrics=["map", "ndcg_cut_10"],
    names=["BM25", "BM25 >> ColBERT", "Dense ColBERT"]
)

ColBERT PRF

You can use ColBERTFactory to obtain ColBERT PRF pipelines, as follows:

colbert_prf_rank = pytcolbert.prf(rerank=False)
colbert_prf_rerank = pytcolbert.prf(rerank=True)

ColBERT PRF requires the ColBERT index to have aligned token ids. During indexing, use the ids=True kwarg for ColBERTIndexer, as follows:

indexer = ColBERTIndexer("/path/to/checkpoint.dnn", "/path/to/index", "index_name", ids=True)

If you use ColBERT PRF in your research, you must cite our ICTIR 2021 paper (citation included below).

All of our results files are available from the paper's Virtual Appendix.

Approximate ANN Scoring and Query Embedding Pruning

This repository contains code to apply the techniques of query embedding pruning [Tonellotto21] and approximate ANN ranking [Macdonald21a].

Query Emebdding pruning can be applied using the following pipeline:

qep_pipe5 = (factory.query_encoder() 
            >> pyterrier_colbert.pruning.query_embedding_pruning(factory, 5) 
            >> factory.set_retrieve(query_encoded=True)
            >> factory.index_scorer(query_encoded=False)
)

where 5 is the number of query embeddings based on collection frquency to retain.

Approximate ANN scoring can be applied using the following pipeline:

ann_pipe = (factory.ann_retrieve_score() % 200) >> factory.index_scorer(query_encoded=True)

where 200 is the number of top-scored ANN candidates to forward for exact scoring.

Demos

• vaswani.ipynb - [Github] [Colab] - demonstrates end-to-end dense retrieval and indexing on the Vaswani corpus (~11k documents)
• colbertprf-msmarco-passages.ipynb - [Github] - demonstrates ColBERT PRF on the TREC Deep Learning track (MSMARCO) passage ranking tasks.
• cikm2021-demos.ipynb - [Github] - demonstrates ANN scoring and Query Embedding Pruning on the TREC Deep Learning track (MSMARCO) passage ranking tasks.
• colbert_text_and_explain.ipynb - [Github] [Colab] - demonstrates using a ColBERT model for scoring text, and for explaining an interaction. If you use one of these interaction diagrams, please cite [Macdonald21].

Resource Requirements

You will need a GPU to use this. Preferable more than one. You will also need lots of RAM - ColBERT requires you load the entire index into memory.

Name	Corpus size	Indexing Time	Index Size
Vaswani	11k abstracts	2 minutes (1 GPU)	163 MB
MSMARCO Passages	8M passages	~24 hours (1 GPU)	192 GB

Installation

This package can be installed using Pip, and then used with PyTerrier. See also the examples notebooks.

pip install -q git+https://github.com/terrierteam/pyterrier_colbert.git
conda install -c pytorch faiss-gpu=1.6.5 # or faiss-cpu
#on Colab: pip install faiss-gpu==1.6.5

NB: ColBERT requires FAISS, namely the faiss-gpu package, to be installed. pip install faiss-gpu does NOT usually work. FAISS recommends using Anaconda to install faiss-gpu. On Colab, you need to resort to pip install. We recommend faiss-gpu version 1.6.3, not 1.7.0.

References

• [Khattab20]: Omar Khattab, Matei Zaharia. ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT. In Proceedings of SIGIR 2020. https://arxiv.org/abs/2004.12832
• [Macdonald20]: Craig Macdonald, Nicola Tonellotto. Declarative Experimentation in Information Retrieval using PyTerrier. Craig Macdonald and Nicola Tonellotto. In Proceedings of ICTIR 2020. https://arxiv.org/abs/2007.14271
• [Macdonald21]: On Single and Multiple Representations in Dense Passage Retrieval. Craig Macdonald, Nicola Tonellotto and Iadh Ounis. In Proceedings of IIR 2021. https://arxiv.org/abs/2108.06279
• [Macdonald21a]: On Approximate Nearest Neighbour Selection for Multi-Stage Dense Retrieval. Craig Macdonald and Nicola Tonellotto. In Proceedings of CIKM 2021. https://arxiv.org/abs/2108.11480
• [Tonellotto21]: Query Embedding Pruning for Dense Retrieval Nicola Tonellotto and Craig Macdonald. In Proceedings of CIKM 2021. https://arxiv.org/abs/2108.10341
• [Wang21]: Xiao Wang, Craig Macdonald, Nicola Tonellotto, Iadh Ounis. Pseudo-Relevance Feedback for Multiple Representation Dense Retrieval. In Proceedings of ICTIR 2021. https://arxiv.org/abs/2106.11251

Credits

• Craig Macdonald, University of Glasgow
• Nicola Tonellotto, University of Pisa
• Sanjana Karumuri, University of Glasgow
• Xiao Wang, University of Glasgow
• Muhammad Hammad Khan, University of Glasgow
• Sean MacAvaney, University of Glasgow
• Sasha Petrov, University of Glasgow