Skip to content

Latest commit

 

History

History
253 lines (186 loc) · 8.07 KB

README.md

File metadata and controls

253 lines (186 loc) · 8.07 KB
Raw

PyTerrier PISA

PyTerrier bindings for the PISA search engine.

Interactive Colab Demo: Open In Colab

Getting Started

These bindings are only available for cpython 3.8-3.11 on manylinux2010_x86_64 platforms. They can be installed via pip:

pip install pyterrier_pisa

Indexing

You can easily index corpora from PyTerrier datasets:

import pyterrier as pt
from pyterrier_pisa import PisaIndex

# from a dataset
dataset = pt.get_dataset('irds:msmarco-passage')
index = PisaIndex('./msmarco-passage-pisa')
index.index(dataset.get_corpus_iter())

You can also select which text field(s) to index. If not specified, all fields of type str will be indexed.

dataset = pt.get_dataset('irds:cord19')
index = PisaIndex('./cord19-pisa', text_field=['title', 'abstract'])
index.index(dataset.get_corpus_iter())

PisaIndex accepts various other options to configure the indexing process. Most notable are:

  • stemmer: Which stemmer to use? Options: porter2 (default), krovetz, none
  • threads: How many threads to use for indexing? Default: 8
  • index_encoding: Which index encoding to use. Default: block_simdbp
  • stops: Which set of stopwords to use. Default: terrier.
# E.g.,
index = PisaIndex('./cord19-pisa', stemmer='krovetz', threads=32)

For some collections you can download pre-built indices from data.terrier.org. PISA indices are prefixed with pisa_.

index = PisaIndex.from_dataset('trec-covid')

Retrieval

From an index, you can build retrieval transformers:

dph = index.dph()
bm25 = index.bm25(k1=1.2, b=0.4)
pl2 = index.pl2(c=1.0)
qld = index.qld(mu=1000.)

These retrievers support all the typical pipeline operations.

Search:

bm25.search('covid symptoms')
#     qid           query     docno     score
# 0     1  covid symptoms  a6avr09j  6.273450
# 1     1  covid symptoms  hdxs9dgu  6.272374
# 2     1  covid symptoms  zxq7dl9t  6.272374
# ..   ..             ...       ...       ...
# 999   1  covid symptoms  m8wggdc7  4.690651

Batch retrieval:

print(dph(dataset.get_topics('title')))
#       qid                     query     docno     score
# 0       1        coronavirus origin  8ccl9aui  9.329109
# 1       1        coronavirus origin  es7q6c90  9.260190
# 2       1        coronavirus origin  8l411r1w  8.862670
# ...    ..                       ...       ...       ...
# 49999  50  mrna vaccine coronavirus  eyitkr3s  5.610429

Experiment:

from pyterrier.measures import *
pt.Experiment(
  [dph, bm25, pl2, qld],
  dataset.get_topics('title'),
  dataset.get_qrels(),
  [nDCG@10, P@5, P(rel=2)@5, 'mrt'],
  names=['dph', 'bm25', 'pl2', 'qld']
)
#    name   nDCG@10    P@5  P(rel=2)@5       mrt
# 0   dph  0.623450  0.720       0.548  1.101846
# 1  bm25  0.624923  0.728       0.572  0.880318
# 2   pl2  0.536506  0.632       0.456  1.123883
# 3   qld  0.570032  0.676       0.504  0.974924

You can also build a retrieval transformer from PisaRetrieve:

from pyterrier_pisa import PisaRetrieve
# from index path:
bm25 = PisaRetrieve('./cord19-pisa', scorer='bm25', bm25_k1=1.2, bm25_b=0.4)
# from dataset
bm25 = PisaRetrieve.from_dataset('trec-covid', 'pisa_unstemmed', scorer='bm25', bm25_k1=1.2, bm25_b=0.4)

Extras

You can access PISA's tokenizer and stemmers using the tokenize function:

import pyterrier_pisa
pyterrier_pisa.tokenize('hello worlds!')
# ['hello', 'worlds']
pyterrier_pisa.tokenize('hello worlds!', stemmer='porter2')
# ['hello', 'world']

FAQ

What retrieval functions are supported?

  • "dph". Parameters: (none)
  • "bm25". Parameters: k1, b
  • "pl2". Parameters: c
  • "qld". Parameters: mu

How do I index [some other type of data]?

PisaIndex accepts an iterator over dicts, each of which containing a docno field and a text field. All you need to do is coerce the data into that format and you're set.

Examples:

# any iterator
def iter_docs():
  for i in range(100):
    yield {'docno': str(i), 'text': f'document {i}'}
index = PisaIndex('./dummy-pisa')
index.index(iter_docs())

# from a dataframe
import pandas as pd
docs = pd.DataFrame([
  ('1', 'test doc'),
  ('2', 'another doc'),
], columns=['docno', 'text'])
index = PisaIndex('./dummy-pisa-2')
index.index(docs.to_dict(orient="records"))

Can I build a doc2query index?

You can use PisaIndex with any document rewriter, such as doc2query or DeepCT. All you need to do is build an indexing pipeline. For example:

pip install --upgrade git+https://github.com/terrierteam/pyterrier_doc2query.git
wget https://git.uwaterloo.ca/jimmylin/doc2query-data/raw/master/T5-passage/t5-base.zip
unzip t5-base.zip
doc2query = Doc2Query(out_attr="exp_terms", batch_size=8)
dataset = pt.get_dataset('irds:vaswani')
index = PisaIndex('./vaswani-doc2query-pisa')
index_pipeline = doc2query >> pt.apply.text(lambda r: f'{r["text"]} {r["exp_terms"]}') >> index
index_pipeline.index(dataset.get_corpus_iter())

Can I build a learned sparse retrieval (e.g., SPLADE) index?

Yes! Example:

import pyt_splade
splade = pyt_splade.Splade()
dataset = pt.get_dataset('irds:msmarco-passage')
index = PisaIndex('./msmarco-passage-splade', stemmer='none')

# indexing
idx_pipeline = splade.doc_encoder() >> index.toks_indexer()
idx_pipeline.index(dataset.get_corpus_iter())

# retrieval

retr_pipeline = splade.query_encoder() >> index.quantized()

msmarco-passage/trec-dl-2019 effectiveness for naver/splade-cocondenser-ensembledistil:

System nDCG@10 R(rel=2)@1000
PISA 0.731 0.872
From Paper 0.732 0.875

What are the supported index encodings and query algorithms?

Right now we support the following index encodings: ef, single, pefuniform, pefopt, block_optpfor, block_varintg8iu, block_streamvbyte, block_maskedvbyte, block_interpolative, block_qmx, block_varintgb, block_simple8b, block_simple16, block_simdbp.

Index encodings are supplied when a PisaIndex is constructed:

index = PisaIndex('./cord19-pisa', index_encoding='ef')

We support the following query algorithms: wand, block_max_wand, block_max_maxscore, block_max_ranked_and, ranked_and, ranked_or, maxscore.

Query algorithms are supplied when you construct a retrieval transformer:

index.bm25(query_algorithm='ranked_and')

Can I import/export from CIFF?

Yes! Using .from_ciff(ciff_file, index_path) and .to_ciff(ciff_file)

# from a CIFF export:
index = PisaIndex.from_ciff('path/to/something.ciff', 'path/to/index.pisa', stemmer='krovetz') # stemmer is optional
# to a CIFF export:
index.to_ciff('path/to/something.ciff')

Note that you need to be careful to set stemmer to match whatever was used when constructing the index; CIFF does not directly store which stemmer was used when building the index. If it's a stemmer that's not supported by PISA, you can set stemmer='none' and apply stemming in a PyTerrier pipeline.

References

Credits

  • Sean MacAvaney, University of Glasgow
  • Craig Macdonald, University of Glasgow