yoh/modular-sentence-encoders-paraphrase

Multi-parallel paraphrase corpus (23 languages)

The contrastive training data of Modular Sentence Encoders: Separating Language Specialization from Cross-Lingual Alignment (ACL 2025). Five English paraphrase datasets, each translated into 22 further languages, published so that the paper's sentence-encoder and alignment stages can be reproduced without re-running the translation.

Most of this corpus is machine-translated. Only the English columns are original human-written text; see Provenance for exactly which columns came from where.

The corpus is multi-parallel: row i of every column is the same content in a different language. That makes it usable three ways, all of which the paper needs:

• monolingual paraphrase pairs — anchor_de / positive_de;
• cross-lingual paraphrase pairs — anchor_en / positive_de;
• translation (parallel) pairs — anchor_en / anchor_de.

Usage

Each dataset is a single Parquet file with flat columns, so a run that needs two languages reads two languages rather than 23:

from datasets import load_dataset

# All 23 languages of one dataset
ds = load_dataset("yoh/modular-sentence-encoders-paraphrase", "quora", split="train")

# Just the columns you need (much faster, much less memory)
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "yoh/modular-sentence-encoders-paraphrase",
    "quora.parquet",
    repo_type="dataset",
)
table = pq.read_table(path, columns=["anchor_en", "positive_de", "negative_de"])

Schema

Every dataset has one column per (role, language) pair, named {role}_{lang}:

role	meaning
anchor	the first sentence of the pair
positive	a paraphrase, entailment or duplicate of the anchor
negative	a hard negative; only in mnli and quora

Language codes are the paper's two-letter codes: am, ar, az, cs, de, en, es, fr, ha, it, kk, ko, ky, mr, nl, pl, ru, rw, te, tr, ug, uz, zh.

config	rows	columns per language	English source
mnli	128,082	anchor / positive / negative	MultiNLI premise, its entailed and its contradicting hypothesis
sentence_compression	179,905	anchor / positive	a news sentence and its compression
simplewiki	102,144	anchor / positive	an English Wikipedia sentence and its Simple English Wikipedia counterpart
altlex	112,610	anchor / positive	an aligned Wikipedia / Simple Wikipedia sentence pair from the AltLex corpus
quora	103,662	anchor / positive / negative	a Quora question, a duplicate of it, and a non-duplicate

626,403 rows in total. 263 rows of the original 626,666 are absent: the translation came back empty in at least one language, and a row is only useful if every language has it.

Provenance

The English side of every dataset is the original, untranslated source data. No non-English column is human-written:

config	non-English columns
mnli	ar, de, es, fr, ru, tr, zh come from XNLI's own translations of the MultiNLI training set; the remaining 15 languages were translated with NLLB-200-3.3B
all others	translated with NLLB-200-3.3B

XNLI's training-set translations are themselves machine translations, produced by the XNLI authors; only XNLI's dev and test sets are human-translated, and they are not used here. So the corpus is machine-translated throughout, with the mnli config mixing two different machine translation systems.

Translation quality was not filtered or scored. The paper's point is precisely that machine-translated data at this scale is good enough to train sentence encoders; Appendix D of the paper reports what that costs on each task.

Licensing

Machine translation. NLLB-200 is released under CC BY-NC 4.0, and its outputs inherit that license. XNLI is likewise non-commercial. The corpus as a whole is therefore published under CC BY-NC 4.0 and is for research use only.

Underlying English data, each with its own terms, which continue to apply:

config	source	terms
mnli	MultiNLI (Williams et al., 2018)	mostly the OANC license (free use, modification, redistribution); some fiction under CC BY 3.0 / CC BY-SA 3.0 or public domain
sentence_compression	google-research-datasets/sentence-compression (Filippova & Altun, 2013)	no license stated by the publisher; distributed "as is"
simplewiki	English and Simple English Wikipedia	CC BY-SA
altlex	AltLex (Hidey & McKeown, 2016)	derived from Wikipedia, CC BY-SA
quora	Quora Question Pairs	Quora's terms for the released dataset

Note the tension in simplewiki and altlex: their source text is share-alike, while the translations carry a non-commercial restriction, and the two cannot be combined cleanly. They are published here under the stricter of the two terms. If you need those two datasets under share-alike terms, translate the English columns yourself with a permissively licensed system.

If you are a rights holder for any of the above and object to this redistribution, please open a discussion on this repository and it will be removed.

Citation

@inproceedings{huang-etal-2025-modular,
    title = "Modular Sentence Encoders: Separating Language Specialization from Cross-Lingual Alignment",
    author = "Huang, Yongxin and Wang, Kexin and Glava{\v{s}}, Goran and Gurevych, Iryna",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    year = "2025",
    pages = "2167--2187",
    url = "https://aclanthology.org/2025.acl-long.108/",
}

Please also cite the source datasets you use and, for the translations, NLLB.

Code: https://github.com/UKPLab/acl2025-modular-sentence-encoders