clips/beir-nl-webis-touche2020

Dataset Card for BEIR-NL Benchmark

Dataset Summary

BEIR-NL is a Dutch-translated version of the BEIR benchmark, a diverse and heterogeneous collection of datasets covering various domains from biomedical and financial texts to general web content. BEIR-NL contains the following tasks:

• Fact-checking: FEVER, Climate-FEVER, SciFact
• Question-Answering: NQ, HotpotQA, FiQA-2018
• Bio-Medical IR: TREC-COVID, NFCorpus
• Argument Retrieval: Touche-2020, ArguAna
• Duplicate Question Retrieval: Quora, CqaDupstack
• Citation-Prediction: SCIDOCS
• Entity Retrieval: DBPedia

Languages

Dutch

Dataset Structure

BEIR-NL adheres to the structure of the original BEIR benchmark. All BEIR datasets must contain a corpus, queries and qrels (relevance judgments file). Qrels can be found in the BEIR repository on GitHub or Hugging Face. They format:

• corpus file: a .jsonl file (jsonlines) that contains a list of dictionaries, each with three fields _id with unique document identifier, title with document title (optional) and text with document paragraph or passage. For example: {"_id": "doc1", "title": "Albert Einstein", "text": "Albert Einstein was een in Duitsland geboren..."}
• queries file: a .jsonl file (jsonlines) that contains a list of dictionaries, each with two fields _id with unique query identifier and text with query text. For example: {"_id": "q1", "text": "Wie ontwikkelde de massa-energie-equivalentieformule?"}
• qrels file: a .tsv file (tab-seperated) that contains three columns, i.e. the query-id, corpus-id and score in this order. Keep 1st row as header. For example: q1 doc1 1

Data Instances

A high level example of any beir dataset:

corpus = {
    "doc1": {
        "title": "Albert Einstein", 
        "text": (
            "Albert Einstein was een in Duitsland geboren theoretisch natuurkundige die de relativiteitstheorie ontwikkelde, "
            "een van de twee pijlers van de moderne natuurkunde (samen met de kwantummechanica). Zijn werk staat ook bekend "
            "om zijn invloed op de wetenschapfilosofie. Hij is bij het grote publiek vooral bekend vanwege zijn massa-energie- "
            "equivalentieformule E = mc^2, die 's werelds beroemdste vergelijking' wordt genoemd. Hij ontving in 1921 de Nobelprijs "
            "voor de Natuurkunde 'voor zijn verdiensten aan de theoretische natuurkunde, en in het bijzonder voor zijn ontdekking "
            "van de wet van het foto-elektrisch effect', een cruciale stap in de ontwikkeling van de kwantumtheorie."
        ),
    },
    "doc2": {
        "title": "", 
        "text": (
            "Tarwebier is een bovengistend bier dat wordt gebrouwen met een groot aandeel tarwe ten opzichte van de hoeveelheid "
            "gemoute gerst. De twee belangrijkste soorten zijn Duits Weißbier en Belgisch witbier; andere soorten zijn onder andere "
            "Lambiek (gemaakt met wilde gist), Berliner Weisse (een troebel, zuur bier) en Gose (een zuur, zout bier)."
        ),
    },
}

queries = {
    "q1": "Wie ontwikkelde de massa-energie-equivalentieformule?",
    "q2": "Welk bier wordt gebrouwen met een groot aandeel tarwe?"
}

qrels = {
    "q1": {"doc1": 1},
    "q2": {"doc2": 1},
}

Data Fields

Examples from all configurations have the following features:

Corpus

• corpus: a dict feature representing the document title and passage text, made up of:
  • _id: a string feature representing the unique document id
    • title: a string feature, denoting the title of the document.
    • text: a string feature, denoting the text of the document.

Queries

• queries: a dict feature representing the query, made up of:
  • _id: a string feature representing the unique query id
  • text: a string feature, denoting the text of the query.

Qrels

• qrels: a dict feature representing the query document relevance judgements, made up of:
  • _id: a string feature representing the query id
    • _id: a string feature, denoting the document id.
    • score: a int32 feature, denoting the relevance judgement between query and document.

Data Splits

Dataset	Website	BEIR-Name	Type	Queries	Corpus	Rel D/Q	BEIR	BEIR-NL
TREC-COVID	Homepage	trec-covid	test	50	171K	493.5	Link	Link
NFCorpus	Homepage	nfcorpus	train dev test	323	3.6K	38.2	Link	Link
NQ	Homepage	nq	train test	3,452	2.68M	1.2	Link	Link
HotpotQA	Homepage	hotpotqa	train dev test	7,405	5.23M	2.0	Link	Link
FiQA-2018	Homepage	fiqa	train dev test	648	57K	2.6	Link	Link
ArguAna	Homepage	arguana	test	1,406	8.67K	1.0	Link	Link
Touche-2020	Homepage	webis-touche2020	test	49	382K	19.0	Link	Link
CQADupstack	Homepage	cqadupstack	test	13,145	457K	1.4	Link	Link
Quora	Homepage	quora	dev test	10,000	523K	1.6	Link	Link
DBPedia	Homepage	dbpedia-entity	dev test	400	4.63M	38.2	Link	Link
SCIDOCS	Homepage	scidocs	test	1,000	25K	4.9	Link	Link
FEVER	Homepage	fever	train dev test	6,666	5.42M	1.2	Link	Link
Climate-FEVER	Homepage	climate-fever	test	1,535	5.42M	3.0	Link	Link
SciFact	Homepage	scifact	train test	300	5K	1.1	Link	Link

Dataset Creation

Curation Rationale

Zero-shot evaluation of information retrieval (IR) models is often performed using BEIR; a large and heterogeneous benchmark composed of multiple datasets, covering different retrieval tasks across various domains. Although BEIR has become a standard benchmark for the zero-shot setup, its exclusively English content reduces its utility for underrepresented languages in IR, including Dutch. To address this limitation and encourage the development of Dutch IR models, we introduce BEIR-NL by automatically translating the publicly accessible BEIR datasets into Dutch.

Source Data

BEIR repository on GitHub.

Annotations

We prompted Gemini-1.5-flash to translate BEIR into Dutch. A small portion of translations were done using GPT-4o-mini and Google Translate, as Gemini declined to translate certain content and had occasional issues with tags in prompts.

Considerations for Using the Data

Other Known Limitations

Not Native Dutch Resources. While BEIR-NL provides a benchmark for evaluating IR models in Dutch, it relies on translations from the original BEIR, which is exclusively in English. This lack of native Dutch datasets limits the ability of BEIR-NL to fully represent and reflect the linguistic nuances and cultural context of the language, and therefore, the complexities of Dutch IR, especially in domain-specific contexts with local terminology and knowledge.

Data Contamination. Many modern IR models are trained on massive corpora that might include content from BEIR. This can result in inflated performances --as models might have already seen the relevant data during different phases of training-- raising concerns about the validity of zero-shot evaluations. Ensuring a truly zero-shot evaluation is a difficult challenge, as many IR models lack transparency regarding the exact composition of training corpora.

Benchmark Validity Over Time. BEIR has become a standard benchmark to evaluate the performance of IR models, attracting a large number of evaluations over time. This extensive usage introduces the risk of overfitting, as researchers might unintentionally train models tailored to perform well on BEIR rather than on broader IR tasks. In addition, advances in IR models and evaluation needs might outpace the benchmark, making it less representative and less relevant. As a result, the relevance and validity of BEIR as well as BEIR-NL may diminish over time.

Additional Information

Licensing Information

This subset (Touche-2020) of BEIR-NL is licensed under the CC BY-NC-SA 4.0 license.

Citation Information

If you find BEIR-NL useful in your research, please consider citing it, as well as the original BEIR benchmark it is derived from:

@misc{banar2024beirnlzeroshotinformationretrieval,
    title={BEIR-NL: Zero-shot Information Retrieval Benchmark for the Dutch Language}, 
     author={Nikolay Banar and Ehsan Lotfi and Walter Daelemans},
     year={2024},
     eprint={2412.08329},
     archivePrefix={arXiv},
     primaryClass={cs.CL},
     url={https://arxiv.org/abs/2412.08329}, 
}

@inproceedings{thakur2021beir,
    title={{BEIR}: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
    author={Nandan Thakur and Nils Reimers and Andreas R{\"u}ckl{\'e} and Abhishek Srivastava and Iryna Gurevych},
    booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
    year={2021},
    url={https://openreview.net/forum?id=wCu6T5xFjeJ}
}