PyTerrier
A Python API for Terrier - v.0.9
Installation
The easiest way to get started with PyTerrier is to use one of our Colab notebooks - look for the 
Linux or Google Colab or Windows
pip install python-terrier- You may need to set JAVA_HOME environment variable if Pyjnius cannot find your Java installation.
macOS
- You need to hava Java installed. Pyjnius/PyTerrier will pick up the location automatically.
pip install python-terrier
Indexing
PyTerrier has a number of useful classes for creating indices:
- You can create an index from TREC formatted collection using TRECCollectionIndexer.
- For TXT, PDF, Microsoft Word files, etc files you can use FilesIndexer.
- For Pandas Dataframe you can use DFIndexer.
- For any abitrary iterable dictionaries, you can use IterDictIndexer.
See the indexing documentation, or the examples in the indexing notebook
Retrieval and Evaluation
topics = pt.io.read_topics(topicsFile)
qrels = pt.io.read_qrels(qrelsFile)
BM25_br = pt.BatchRetrieve(index, wmodel="BM25")
res = BM25_br.transform(topics)
pt.Utils.evaluate(res, qrels, metrics = ['map'])See also the retrieval documentation, or the worked example in the retrieval and evaluation notebook
Experiment - Perform Retrieval and Evaluation with a single function
PyTerrier provides an Experiment function, which allows to compare multiple retrieval approaches on the same queries & relevance assessments:
pt.Experiment([BM25_br, PL2_br], topics, qrels, ["map", "ndcg"])There is a worked example in the experiment notebook
Pipelines
PyTerrier makes it easy to develop complex retrieval pipelines using Python operators such as >> to chain different retrieval components. Each retrieval approach is a transformer, having one key method, transform(), which takes a single Pandas dataframe as input, and returns another dataframe. Two examples might encapsulate applying the sequential dependence model, or a query expansion process:
sdm_bm25 = pt.rewrite.SDM() >> pt.BatchRetrieve(indexref, wmodel="BM25")
bo1_qe = BM25_br >> pt.rewrite.Bo1QueryExpansion() >> BM25_brThere is documentation on transformer operators as well as example pipelines show other common use cases. For more information, see the PyTerrier data model.
Neural Reranking and Dense Retrieval
PyTerrier has additional plugins for BERT (through OpenNIR), T5, ColBERT, ANCE, DeepCT and doc2query.
- OpenNIR: [Github] [Documentation]
- PyTerrier_ANCE: [Github] - dense retrieval
- PyTerrier_ColBERT: [Github] - dense retrieval and/or neural reranking
- PyTerrier_PISA: [Github] - fast in-memory indexing and retrieval using PISA
- PyTerrier_T5: [Github] - neural reranking: monoT5, duoT5
- PyTerrier_doc2query: [Github] - neural augmented indexing
- PyTerrier_SPLADE: [Github] - neural augmented indexing
- PyTerrier_DeepCT: [Github] - neural augmented indexing
You can see examples of how to use these, including notebooks that run on Google Colab, in the contents of our ECIR 2021 tutorial.
Learning to Rank
Complex learning to rank pipelines, including for learning-to-rank, can be constructed using PyTerrier's operator language. For example, to combine two features and make them available for learning, we can use the ** operator.
two_features = BM25_br >> (
pt.BatchRetrieve(indexref, wmodel="DirichletLM") **
pt.BatchRetrieve(indexref, wmodel="PL2")
)See also the learning to rank documentation, as well as the worked examples in the learning-to-rank notebook
Dataset API
PyTerrier allows simple access to standard information retrieval test collections through its dataset API, which can download the topics, qrels, corpus or, for some test collections, a ready-made Terrier index.
topics = pt.get_dataset("trec-robust-2004").get_topics()
qrels = pt.get_dataset("trec-robust-2004").get_qrels()
pt.Experiment([BM25_br, PL2_br], topics, qrels, eval_metrics)You can index datasets that include a corpus using IterDictIndexer and get_corpus_iter:
dataset = pt.get_dataset('irds:cord19/trec-covid')
indexer = pt.IterDictIndexer('./cord19-index')
index_ref = indexer.index(dataset.get_corpus_iter(), fields=('title', 'abstract'))You can use pt.list_datasets() to see available test collections - if your favourite test collection is missing, you can submit a Pull Request.
All datasets from the ir_datasets package are available
under the irds: prefix. E.g., use pt.datasets.get_dataset("irds:medline/2004/trec-genomics-2004")
to get the TREC Genomics 2004 dataset. A full catalogue of ir_datasets is available here.
Index API
All of the standard Terrier Index API can be access easily from Pyterrier.
For instance, accessing term statistics is a single call on an index:
index.getLexicon()["circuit"].getDocumentFrequency()There are lots of examples in the index API notebook
Documentation
More documentation for PyTerrier is available at https://pyterrier.readthedocs.io/en/latest/.
Open Source Licence
PyTerrier is subject to the terms detailed in the Mozilla Public License Version 2.0. The Mozilla Public License can be found in the file LICENSE.txt. By using this software, you have agreed to the licence.
Citation Licence
The source and binary forms of PyTerrier are subject to the following citation license:
By downloading and using PyTerrier, you agree to cite at the undernoted paper describing PyTerrier in any kind of material you produce where PyTerrier was used to conduct search or experimentation, whether be it a research paper, dissertation, article, poster, presentation, or documentation. By using this software, you have agreed to the citation licence.
@inproceedings{pyterrier2020ictir,
author = {Craig Macdonald and Nicola Tonellotto},
title = {Declarative Experimentation inInformation Retrieval using PyTerrier},
booktitle = {Proceedings of ICTIR 2020},
year = {2020}
}
Credits
- Alex Tsolov, University of Glasgow
- Craig Macdonald, University of Glasgow
- Nicola Tonellotto, University of Pisa
- Arthur Cรขmara, TU Delft
- Alberto Ueda, Federal University of Minas Gerais
- Sean MacAvaney, Georgetown University/University of Glasgow
- Chentao Xu, University of Glasgow
- Sarawoot Kongyoung, University of Glasgow
- Zhan Su, Copenhagen University
- Marcus Schutte, TU Delft