EViRAL: Ede-Vietnamese Retrieval Across Languages

EViRAL is a cross-lingual information retrieval benchmark for Ede (ISO 639-3: rad, Glottocode: rade1241), a low-resource Austronesian language spoken primarily in the Central Highlands of Vietnam. The dataset pairs Ede queries with Vietnamese passages sourced from WebFAQ, forming the first IR benchmark for this language.

Queries are derived from the Vietnamese subset of WebFAQ and translated into Ede using the NIRVLab/ViEde machine translation model (Vietnamese-to-Ede). The corpus and relevance judgments (qrels) are retained from the original WebFAQ Vietnamese retrieval subset without modification.

Dataset Structure

The dataset contains three subsets following the standard BEIR/MTEB format:

queries — questions in both Ede and Vietnamese, split into train, validation, and test.

Column	Type	Description
query_id	string	Unique query identifier
query_vi	string	Original query in Vietnamese
query_ede	string	Machine-translated query in Ede

corpus — Vietnamese FAQ passages.

Column	Type	Description
corpus_id	string	Unique passage identifier
title	string	Passage title (may be empty)
passage	string	Passage text in Vietnamese

qrels — relevance judgments mapping queries to relevant passages.

Column	Type	Description
query_id	string	Query identifier
corpus_id	string	Relevant passage identifier
score	float	Relevance score (1 = relevant)

Splits

Split	Queries	Qrels
train	~79,700	~79,700
validation	~17,100	~17,100
test	~17,100	~17,100
corpus	124,000 passages	—

The split was produced using a stratified query-disjoint strategy: each query appears in exactly one split, and stratification is performed by query character-length bucket (short / medium / long) to preserve distributional balance across splits. This design prevents data leakage through shared queries between train and evaluation sets. Split ratio: 70 / 15 / 15.

Loading the Dataset

from datasets import load_dataset

# Load queries (all splits)
queries = load_dataset("NIRVLab/EViRAL", "queries")

# Load corpus
corpus = load_dataset("NIRVLab/EViRAL", "corpus", split="corpus")

# Load qrels (all splits)
qrels = load_dataset("NIRVLab/EViRAL", "qrels")

# Example: access test split
test_queries = queries["test"]
test_qrels   = qrels["test"]

Evaluation with BEIR

from beir import util
from beir.retrieval.evaluation import EvaluateRetrieval

# Load EViRAL subsets into BEIR-compatible format
# corpus: {corpus_id: {"title": ..., "text": ...}}
# queries: {query_id: query_ede_text}
# qrels: {query_id: {corpus_id: score}}

Evaluation with MTEB

import mteb

# EViRAL follows the standard BEIR retrieval task format.
# To evaluate a model:
model = mteb.get_model("your-model-name")
# Pass corpus, queries, and qrels as a custom retrieval task.

Construction Pipeline

1. Source data. Vietnamese queries, corpus, and qrels were loaded from PaDaS-Lab/webfaq-retrieval (vie-queries, vie-corpus, vie-qrels subsets).

2. Quality filtering. Queries shorter than 5 characters and exact/near-duplicate queries (by normalized MD5 hash) were removed.

3. Translation. Surviving queries were translated from Vietnamese to Ede using NIRVLab/ViEde, a fine-tuned mBART-based model (BLEU 22.8, ChrF++ 46.2 on held-out evaluation). Translation was performed with greedy decoding (num_beams=1), max 64 source/target tokens, using float16 precision on GPU.

4. Stratified split. Query IDs were divided into train / validation / test (70/15/15) using stratified shuffle split on query length bucket, ensuring no query appears in more than one split.

5. Integrity check. All qrels were verified to reference only valid query IDs and corpus IDs present in their respective subsets.

Languages

Language	ISO 639-3	Glottocode	Role in dataset
Vietnamese	vie	viet1252	Corpus passages, parallel queries
Ede (Rade)	rad	rade1241	Primary query language

Ede (also written Ede, Rhade, or Rade) is an Austronesian language of the Malayo-Polynesian branch spoken by approximately 400,000 people in the Central Highlands of Vietnam, primarily in Dak Lak and Phu Yen provinces.

Intended Use

EViRAL is intended for:

• Benchmarking cross-lingual dense and sparse retrieval models on a low-resource language pair.
• Training and evaluating multilingual text embedding models for underrepresented languages.
• Studying the effectiveness of machine-translated queries in cross-lingual IR evaluation.
• Supporting NLP research for Ede and other Central Highlands minority languages of Vietnam.

Limitations

Machine-translated queries. Ede queries were produced by NIRVLab/ViEde without human post-editing. Translation quality (BLEU 22.8, ChrF++ 46.2) is moderate; errors may propagate to retrieval evaluation scores. Results should be interpreted with this in mind and compared against human-translated or natively authored baselines when available.

Single relevant passage per query. Each query is paired with exactly one relevant passage (the answer from the originating FAQ page). Additional relevant passages on other websites are not labeled. This sparse relevance structure is shared with the original WebFAQ retrieval subsets.

No guarantee of factual accuracy. Passages reflect content from public FAQ pages at crawl time (Common Crawl 2022-2024). They may contain outdated, biased, or incorrect information.

Language detection noise. A small proportion of Vietnamese FAQ pairs may contain mixed-language content or brand names that escaped language detection in the original WebFAQ pipeline.

Translationese effects. Cross-lingual IR benchmarks built via machine translation may exhibit translationese artifacts, which can introduce bias toward translation-aware retrieval methods. See Artetxe et al. (2020) and related work for discussion.

Ethical Considerations

Data provenance. The corpus is derived from public FAQ pages collected by Common Crawl and processed by the Web Data Commons project. Content was published by website owners with the intent of public access. Downstream users should verify compliance with the terms of use of source websites and Common Crawl.

Low-resource language representation. Ede is a seriously under-resourced language in NLP. This dataset is intended to contribute to, not substitute for, community-led efforts to develop language resources for Ede speakers. The authors acknowledge that machine-translated queries do not reflect native speaker usage and encourage future work involving native Ede speakers in data creation and evaluation.

No personally identifiable information. The dataset does not contain personally identifiable information. All content is sourced from publicly accessible FAQ pages.

License compliance. The source dataset PaDaS-Lab/webfaq-retrieval is released under CC BY 4.0. EViRAL inherits this license. Users are required to attribute the original WebFAQ dataset and this dataset when publishing results.

Citation

If you use EViRAL in your research, please cite this dataset and the following works:

WebFAQ (source dataset)

@inproceedings{dinzinger2025webfaq,
  title={WebFAQ: A Multilingual Collection of Natural Q\&A Datasets for Dense Retrieval},
  author={Dinzinger, Michael and Caspari, Laura and Ghosh Dastidar, Kanishka and Mitrovi{\'c}, Jelena and Granitzer, Michael},
  booktitle={Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages={3802--3811},
  year={2025}
}

BEIR benchmark (evaluation framework)

@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 Abhijit Bhole and Iryna Gurevych},
  booktitle = {Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year      = {2021},
  url       = {https://openreview.net/forum?id=wCu6T5xFjeJ}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, consistent with the source WebFAQ retrieval dataset.

Acknowledgments

The corpus and relevance judgments are derived from the WebFAQ project (Dinzinger et al., 2025), which collected FAQ data from Common Crawl and Web Data Commons snapshots. Ede query translation was performed using the NIRVLab/ViEde model developed by the Network for Intelligent Research Vietnam (NIRVLab). The split methodology follows practices established by BEIR (Thakur et al., 2021) and MTEB (Muennighoff et al., 2022).