EXCISE-corpora / README.md
mohammed933's picture
Card: link the DataScienceUIBK/EXCISE repository
cb6923c verified
|
Raw History Blame Contribute Delete
2.97 kB
metadata
license: other
license_name: source-collection-terms
language:
  - en
task_categories:
  - text-retrieval
task_ids:
  - document-retrieval
tags:
  - information-retrieval
  - colbert
  - late-interaction
  - negation
  - exclusion
  - benchmark
size_categories:
  - 1M<n<10M
pretty_name: X-BENCH
configs:
  - config_name: benchmark
    data_files:
      - split: test
        path: benchmark/test.parquet
      - split: train
        path: benchmark/train.parquet
  - config_name: corpus
    data_files:
      - split: excluir
        path: corpus/excluir.parquet
      - split: fiqa
        path: corpus/fiqa.parquet
      - split: trec_covid
        path: corpus/trec_covid.parquet
      - split: esgenius
        path: corpus/esgenius.parquet
      - split: edgar
        path: corpus/edgar.parquet
      - split: eurlex
        path: corpus/eurlex.parquet

X-BENCH

Exclusion queries of the form "X but not Z" over six retrieval collections, released with EXCISE: Query-Side Exclusion for Late-Interaction Retrieval. A query counts as a success@10 only when its gold document is in the top ten and no excluded document is. Code: github.com/DataScienceUIBK/EXCISE.

Subsets

benchmark holds the queries: test (1,860) and train (1,100, ExcluIR and FiQA only). Each query names its gold document (pos_id) and one excluded document per Z (neg_ids). corpus holds the documents, one split per collection, as _id, title and text. A query's documents are found by _id in the corpus split of its collection.

Query ids are unique across both splits. Document ids are unique within each collection, and every id the benchmark cites is in its corpus.

Collection Documents Test queries
excluir 90,406 554 in-domain
fiqa 57,638 208 in-domain
trec_covid 171,332 343 zero-shot
esgenius 995 206 zero-shot
edgar 676,925 353 zero-shot
eurlex 322,938 196 zero-shot

EDGAR and EUR-Lex are split into passages (a median of 227 and 137 words); the other four are kept as their sources publish them.

Tiers

T1 states the exclusion with a cue such as not or excluding (693 test queries), T2 has no cue at all (569), and T3 rules out two topics (598).

Loading

from datasets import load_dataset
test  = load_dataset("mohammed933/EXCISE-corpora", "benchmark", split="test")
edgar = load_dataset("mohammed933/EXCISE-corpora", "corpus", split="edgar")

The code in the GitHub repository reads <collection>/corpus.jsonl instead, and checks each file against its sha256 pin in data/benchmark/CORPUS_REQUIRED.json:

python scripts/download.py --what corpora     # into data/datasets/<collection>/corpus.jsonl
python scripts/download.py --verify

ExcluIR document ids are positional, so the pins are what identify a corpus. The collections are redistributed from ExcluIR, FiQA, TREC-COVID, ESGenius, EDGAR-CORPUS and MultiEURLEX, and each remains under its source's terms.