orgrctera/beir_webis_touche2020

Webis-Touche2020 (BEIR) — Argument retrieval

Dataset description

Touché 2020 was the first CLEF shared task dedicated to argument retrieval: given an information need expressed as a controversial topic or comparison-style question, systems must return argumentative units (not merely factual snippets) that help a user understand positions, evidence, and counterpoints. The Webis lab framed two complementary tasks—retrieval from a focused debate collection (societally relevant controversies) and retrieval from a broad web crawl (comparative, everyday decisions). The slice packaged in BEIR as webis-touche2020 corresponds to the argument retrieval for controversial questions setting, using the args.me-style corpus of mined arguments and the official test topics and judgments from the Touché 2020 evaluation campaign.

BEIR (Benchmarking IR) unified this release with many other public corpora so dense, sparse, and hybrid retrievers can be compared under a single zero-shot evaluation protocol across heterogeneous IR tasks. In BEIR’s table, Touche-2020 (webis-touche2020) is listed as test-only: 49 queries and a corpus on the order of 382K passages, with substantial multi-document relevance per query (high average relevant documents per topic).

This Hub repository (orgrctera/beir_webis_touche2020) exposes the BEIR test split in Parquet form for retrieval pipelines. Each row is one evaluation query with graded relevance judgments (expected_output) pointing at corpus document identifiers aligned with the BEIR webis-touche2020 release.

Task: retrieval (Webis-Touche2020 in BEIR)

The task is ad hoc argument retrieval:

1. Input: a short controversial question or stance-oriented topic (the query), e.g. policy or health debates.
2. Output: a ranked list of argument IDs from the Touché / BEIR corpus, or scores over the collection, such that human-judged relevant arguments—at the grades released in qrels—rank highly.

Evaluation follows standard IR metrics (e.g. nDCG@k, MAP), with graded labels in the upstream Touché setting (retrieval scores in expected_output use integer grades such as 1 and 2, not only binary relevance).

Note: Full retrieval evaluation also requires the corpus (argument texts keyed by the same _ids as in BEIR). This dataset provides the query + qrels side in tabular form; pair it with the BEIR webis-touche2020 corpus from the BEIR distribution or mirrors such as BeIR/webis-touche2020.

Data format (this repository)

Each record includes:

Field	Description
id	UUID for this example row.
input	Natural-language query (controversial topic / question).
expected_output	JSON string: list of objects {"id": "<corpus-doc-id>", "score": <relevance>}. IDs match BEIR corpus passage identifiers; scores encode graded relevance from the Touché qrels.
metadata.query_id	Original BEIR / Touché query identifier (string).
metadata.split	Split name: test (BEIR only distributes the test split for this subset).

Example 1

{
  "id": "7f25e99c-f6be-4053-a183-8424edb5c3b5",
  "input": "Should everyone get a universal basic income?",
  "expected_output": "[{\"id\": \"4d103774-2019-04-18T13:49:55Z-00005-000\", \"score\": 2}, {\"id\": \"4d1037f0-2019-04-18T11:08:29Z-00000-000\", \"score\": 2}, {\"id\": \"74547446-2019-04-18T19:11:29Z-00005-000\", \"score\": 1}, ...]",
  "metadata.query_id": "50",
  "metadata.split": "test"
}

Example 2

{
  "id": "c695813a-320b-430e-9b62-9ed8ee43121f",
  "input": "Is vaping with e-cigarettes safe?",
  "expected_output": "[{\"id\": \"18710bc8-2019-04-18T16:37:00Z-00000-000\", \"score\": 2}, {\"id\": \"18710bc8-2019-04-18T16:37:00Z-00001-000\", \"score\": 2}, {\"id\": \"8e48cbbe-2019-04-18T14:00:01Z-00003-000\", \"score\": 1}, ...]",
  "metadata.query_id": "2",
  "metadata.split": "test"
}

References

Touché 2020 (shared task overview)

Alexander Bondarenko, Maik Fröbe, Meriem Beloucif, Lukas Gienapp, Yamen Ajjour, Alexander Panchenko, Chris Biemann, Benno Stein, Henning Wachsmuth, Martin Potthast, Matthias Hagen
Overview of Touché 2020: Argument Retrieval — Extended abstract, CLEF 2020 (Thessaloniki, Greece).

Abstract (excerpt): “This paper is a condensed report on Touché: the first shared task on argument retrieval that was held at CLEF 2020. With the goal to create a collaborative platform for research in argument retrieval, we run two tasks: (1) supporting individuals in finding arguments on socially important topics and (2) supporting individuals with arguments on everyday personal decisions.”

• Task description (Shared Task 1 — controversial questions): Touché 2020 — Shared Task 1
• Lab site: Touché at CLEF 2020
• Full proceedings paper (Springer LNCS): 10.1007/978-3-030-58219-7_26

BEIR benchmark (Webis-Touche2020 as a subset)

Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, Iryna Gurevych
BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models
NeurIPS 2021 (Datasets and Benchmarks Track).

Abstract (from the BEIR paper): “Existing neural information retrieval (IR) models have often been studied in homogeneous and narrow settings, which has considerably limited insights into their out-of-distribution (OOD) generalization capabilities. To address this, and to facilitate researchers to broadly evaluate the effectiveness of their models, we introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains and evaluate 10 state-of-the-art retrieval systems including lexical, sparse, dense, late-interaction and re-ranking architectures on the BEIR benchmark.”

• Paper: arXiv:2104.08663 — OpenReview
• Code and dataset pointers: BEIR on GitHub

Related resources

• Classic BEIR-style files on Hugging Face: BeIR/webis-touche2020 (corpus / queries / qrels).

Citation

If you use the Touché 2020 task and judgments, cite the Touché overview (Bondarenko et al., CLEF 2020) and follow any license terms of the underlying args.me / Webis collections. If you use the BEIR packaging and evaluation protocol, cite the BEIR paper (Thakur et al., NeurIPS 2021). BibTeX for BEIR is available in the official repository.

License

The Touché overview publication is marked CC BY 4.0 by its authors. Corpus and redistribution terms for the underlying Webis/args.me data should be confirmed against the original Touché / args.me license for your use case; this card uses cc-by-4.0 as a common umbrella where applicable.

---

Dataset card maintained for the orgrctera/beir_webis_touche2020 Hub repository.