macavaney's picture
Update README.md
77868da verified
|
raw
history blame
4.02 kB
metadata
tags:
  - pyterrier
  - pyterrier-artifact
  - pyterrier-artifact.sparse_index
  - pyterrier-artifact.sparse_index.pisa
task_categories:
  - text-retrieval
viewer: false

MS MARCO PISA Index

Description

This is an index of the MS MARCO passage (v1) dataset with PISA. It can be used for passage retrieval using lexical methods.

Usage

>>> from pyterrier_pisa import PisaIndex
>>> index = PisaIndex.from_hf('macavaney/msmarco-passage.pisa')
>>> bm25 = index.bm25()
>>> bm25.search('terrier breeds')
    qid           query    docno      score  rank
0     1  terrier breeds  1406578  22.686367     0
1     1  terrier breeds  5785957  22.611134     1
2     1  terrier breeds  7455374  22.592781     2
3     1  terrier breeds  3984886  22.242958     3
4     1  terrier breeds  3984893  22.009525     4
...

Benchmarks

TREC DL 2019

Code
from ir_measures import nDCG, RR, MAP, R
import pyterrier as pt
from pyterrier_pisa import PisaIndex
index = PisaIndex.from_hf('macavaney/msmarco-passage.pisa')
dataset = pt.get_dataset('irds:msmarco-passage/trec-dl-2019/judged')
pt.Experiment(
  [index.bm25(), index.qld(), index.dph(), index.pl2()],
  dataset.get_topics(),
  dataset.get_qrels(),
  [nDCG@10, nDCG, RR(rel=2), MAP(rel=2), R(rel=2)@1000],
  ['BM25', 'QLD', 'DPH', 'PL2'],
  round=4,
)
name nDCG@10 nDCG RR(rel=2) AP(rel=2) R(rel=2)@1000
0 BM25 0.4989 0.6023 0.6804 0.3031 0.7555
1 QLD 0.468 0.5984 0.6047 0.3037 0.7601
2 DPH 0.4975 0.5907 0.6674 0.3009 0.7436
3 PL2 0.4503 0.5681 0.6495 0.2679 0.7304

TREC DL 2020

Code
from ir_measures import nDCG, RR, MAP, R
import pyterrier as pt
from pyterrier_pisa import PisaIndex
index = PisaIndex.from_hf('macavaney/msmarco-passage.pisa')
dataset = pt.get_dataset('irds:msmarco-passage/trec-dl-2020/judged')
pt.Experiment(
  [index.bm25(), index.qld(), index.dph(), index.pl2()],
  dataset.get_topics(),
  dataset.get_qrels(),
  [nDCG@10, nDCG, RR(rel=2), MAP(rel=2), R(rel=2)@1000],
  ['BM25', 'QLD', 'DPH', 'PL2'],
  round=4,
)
name nDCG@10 nDCG RR(rel=2) AP(rel=2) R(rel=2)@1000
0 BM25 0.4793 0.5963 0.6529 0.2974 0.8048
1 QLD 0.4511 0.587 0.5812 0.2879 0.8125
2 DPH 0.4586 0.5704 0.6123 0.2779 0.798
3 PL2 0.4552 0.5609 0.5788 0.2666 0.7772

MS MARCO Dev (small)

Code
from ir_measures import RR, R
import pyterrier as pt
from pyterrier_pisa import PisaIndex
index = PisaIndex.from_hf('macavaney/msmarco-passage.pisa')
dataset = pt.get_dataset('irds:msmarco-passage/dev/small')
pt.Experiment(
  [index.bm25(), index.qld(), index.dph(), index.pl2()],
  dataset.get_topics(),
  dataset.get_qrels(),
  [RR@10, R@1000],
  ['BM25', 'QLD', 'DPH', 'PL2'],
  round=4,
)
name RR@10 R@1000
0 BM25 0.185 0.8677
1 QLD 0.1683 0.8542
2 DPH 0.1782 0.8605
3 PL2 0.1741 0.8607

Reproduction

>>> import pyterrier_pisa
>>> import pyterrier as pt
>>> idx = pyterrier_pisa.PisaIndex('msmarco-passage.pisa')
>>> idx.indexer().index(pt.get_dataset('irds:msmarco-passage').get_corpus_iter())

Metadata

{
  "type": "sparse_index",
  "format": "pisa",
  "package_hint": "pyterrier-pisa",
  "stemmer": "porter2"
}