eherra/EuroPIRQ-retrieval

EuroPIRQ: European Parallel Information Retrieval Queries

Dataset Details

The EuroPIRQ retrieval dataset is a multilingual collection designed for evaluating retrieval and cross-lingual retrieval tasks. Dataset contains 10,000 parallel passages & 100 parallel queries (synthetic) in three languages: English🇬🇧, Portuguese🇵🇹, and Finnish🇫🇮, constructed from the European Union's DGT-Acquis corpus.

• Languages: English (en), Portuguese (pt), Finnish (fi)
• Domain: Legal / European Union legislation
• Parallel passages: 10,000 per language each having parallel content
• Parallel (synthetic) queries: 100 per language each having parallel content

---

How to evaluate on this task with MTEB

You can evaluate an embedding model on this dataset using the following code:

import mteb

task = mteb.get_task("EuroPIRQRetrieval")
evaluator = mteb.MTEB([task])

model_name = "<YOUR_MODEL_NAME_HERE>"  # e.g., "intfloat/multilingual-e5-small"

model = mteb.get_model(model_name)
evaluator.run(model)

I also added simple retrieval evaluation (MRR@10) example script to examples/faiss/retrieval_evals.py file.

Retrieval Evaluation

Below, you will find the evaluation results of vector search retrieval using FAISS, comparing the Flat and HNSW index types, with 8 different multilingual embedding models. The evaluation also includes results from using quantized embeddings (int8) and binary in retrieval.

With a Radar chart highlighting the quantization performance impact:

Links to evaluated models in HuggingFace:

• intfloat/multilingual-e5-small
• paraphrase-multilingual-MiniLM-L12-v2
• intfloat/multilingual-e5-base
• sentence-transformers/LaBSE
• sentence-transformers/stsb-xlm-r-multilingual
• BAAI/bge-m3
• intfloat/multilingual-e5-large
• intfloat/multilingual-e5-large-instruct

---

Dataset visualization

To better understand the dataset, this figure illustrates the corpus (en_corpus) dataset clustered into 10 thematic groups. The left and right panels show 2D and 3D t-SNE projections. Colors represents distinct clusters, with labels generated by GPT-4o via analysis of 100 representative passage samples per group.

Dataset Structure

Repository has parquet-files in MTEB format e.g:

• en-corpus
• en-qrels
• en-queries

Repository also has /data folder which consists of two separate JSONL files per language, one for queries and one for the full corpus chunks. JSONL files has the following structure:

data/[en|fi|pt]/full_corpus.jsonl

{
    "id": 1,
    "content": "On a proper construction of Article 24 of Commission Regulation (EEC) No 3719/88 of 16 November 1988 laying down common detailed rules for the application ..."
},
{
    "id": 2,
    "content": "The presentation to customs of goods introduced into the Community, in terms of Article 4(19) of Council Regulation (EEC) No 2913/92 of 12 October 1992 establishing ..."
},
// ...more data

data/[en|fi|pt]/queries.jsonl

{
    "id": "71b4463c-0376-44a6-9aeb-597eca7573da",
    "query": "What concerns does the Committee have about the legislative process for a future directive?",
    "context": "The Committee recognises that consulting all the interested parties \u2014 the market, supervisory authorities and governments \u2014 will necessarily ...",
    "chunk_id": "1139"
},
{
    "id": "e0d3ba10-86bf-4dfb-b49d-7545f0eb1c5d",
    "query": "What role did Serge Durande play in the process involving the Statement of Objections sent to various parties in April 2007?",
    "context": "The Statement of Objections was adopted on 24 April 2007 and sent to the following ...",
    "chunk_id": "6686"
},
// ...more data

Data Fields

queries:

• id: A unique identifier for each entry.
• query: A question generated from the context chunk, used for retrieval tasks.
• context: The corresponding passage (or chunk) from the respective language, providing the context for the query.
• chunk_id: The identifier linking the query to its corresponding passage

full_corpus:

• id: A unique identifier for each text chunk, same across all three languages. Maps to chunk_id in queries.jsonl
• content: The actual text content of the chunk.

---

Dataset Construction

This dataset was constructed by processing the DGT-Acquis Paragraph-Level Corpus (da1-pc) through the following steps:

1. Text File Extraction: Paragraph-level text files were downloaded from the DGT-Acquis and merged into a structured format.
2. Text Cleaning: Concatenated words were corrected, and specific tags (e.g., <HT TYPE="SUP">) were removed.
3. Sentence Splitting: Chunks of 70–100 words (English/Portuguese) and 60–100 words (Finnish) were selected.
4. Language Detection: Texts were validated using the lingua-py library to be in correct language.
5. Cosine Similarity: Semantic similarity checks was performed (cosine score ≥ 0.8) to verify the cross-lingual alignment of the parallel triplets across all three languages.
6. Final Selection: From the translation triplets that met all the criteria above, the first 10,000 sections were chosen for the final dataset.
7. Synthetic Query Generation: 100 randomized English chunks were used to generate synthetic query-passage pairs using LlamaIndex and GPT-4o. Synthetic questions were then created from the corresponding Finnish and Portuguese chunks using specific prompts maintaining consistency with the original English questions and context

Final Dataset Composition:

• 3x 10,000 parallel context chunks across the languages
• 3x 100 synthetic questions across the languages, mapped to the corresponding passage.

Source Data

This dataset is derived from the DGT-Acquis corpus (website, & research paper), which is owned by the European Commission and released under the European Union Public License (EUPL).

• Original Corpus: DGT-Acquis (paragraph-level, 2004–2011).
• Producers: The European Commission, Directorate-General for Translation (DGT).
• Data Processing: The dataset was processed by a custom pipeline, which included text extraction, cleaning, chunk selection, language detection, similarity checks, and synthetic question generation.

Citation

If you use this dataset, please cite the dataset as well as mteb, as this dataset likely includes additional processing as a part of the MMTEB Contribution.

@misc{eherra_2025_europirq,
  author = { {Elias Herranen} },
  publisher = { Hugging Face },
  title = { EuroPIRQ: European Parallel Information Retrieval Queries },
  url = { https://huggingface.co/datasets/eherra/EuroPIRQ-retrieval },
  year = {2025},
}


@article{enevoldsen2025mmtebmassivemultilingualtext,
  title={MMTEB: Massive Multilingual Text Embedding Benchmark},
  author={Kenneth Enevoldsen and Isaac Chung and Imene Kerboua and Márton Kardos and Ashwin Mathur and David Stap and Jay Gala and Wissam Siblini and Dominik Krzemiński and Genta Indra Winata and Saba Sturua and Saiteja Utpala and Mathieu Ciancone and Marion Schaeffer and Gabriel Sequeira and Diganta Misra and Shreeya Dhakal and Jonathan Rystrøm and Roman Solomatin and Ömer Çağatan and Akash Kundu and Martin Bernstorff and Shitao Xiao and Akshita Sukhlecha and Bhavish Pahwa and Rafał Poświata and Kranthi Kiran GV and Shawon Ashraf and Daniel Auras and Björn Plüster and Jan Philipp Harries and Loïc Magne and Isabelle Mohr and Mariya Hendriksen and Dawei Zhu and Hippolyte Gisserot-Boukhlef and Tom Aarsen and Jan Kostkan and Konrad Wojtasik and Taemin Lee and Marek Šuppa and Crystina Zhang and Roberta Rocca and Mohammed Hamdy and Andrianos Michail and John Yang and Manuel Faysse and Aleksei Vatolin and Nandan Thakur and Manan Dey and Dipam Vasani and Pranjal Chitale and Simone Tedeschi and Nguyen Tai and Artem Snegirev and Michael Günther and Mengzhou Xia and Weijia Shi and Xing Han Lù and Jordan Clive and Gayatri Krishnakumar and Anna Maksimova and Silvan Wehrli and Maria Tikhonova and Henil Panchal and Aleksandr Abramov and Malte Ostendorff and Zheng Liu and Simon Clematide and Lester James Miranda and Alena Fenogenova and Guangyu Song and Ruqiya Bin Safi and Wen-Ding Li and Alessia Borghini and Federico Cassano and Hongjin Su and Jimmy Lin and Howard Yen and Lasse Hansen and Sara Hooker and Chenghao Xiao and Vaibhav Adlakha and Orion Weller and Siva Reddy and Niklas Muennighoff},
  publisher = {arXiv},
  journal={arXiv preprint arXiv:2502.13595},
  year={2025},
  url={https://arxiv.org/abs/2502.13595},
  doi = {10.48550/arXiv.2502.13595},
}

@article{muennighoff2022mteb,
  author = {Muennighoff, Niklas and Tazi, Nouamane and Magne, Loïc and Reimers, Nils},
  title = {MTEB: Massive Text Embedding Benchmark},
  publisher = {arXiv},
  journal={arXiv preprint arXiv:2210.07316},
  year = {2022}
  url = {https://arxiv.org/abs/2210.07316},
  doi = {10.48550/ARXIV.2210.07316},
}