# Parallel corpora of EU law in 24 languages for MT and multilingual LLMs

> EU legislation is published in all 24 official EU languages, which are equally authentic, so EUR-Lex is the largest source of professionally translated legal text in Europe; ready-made parallel corpora include the Commission's DGT-TM (24 languages, latest release 2021), JRC-Acquis (22 languages) and MultiEURLEX (65k laws in 23 languages), and you can build your own by aligning language versions on the CELEX number and article number.

Source: https://data.arietalabs.com/guides/multilingual-eu-law-parallel-corpus/
Published: 2026-11-13 · Updated: 2026-11-14 · Author: Nejc Fošnarič, Founder, Arieta Labs (https://data.arietalabs.com/about/#nejc-fosnaric/)

EU law is the largest body of professionally translated legal text in Europe, because the European Union publishes its legislation in all 24 official languages and treats every language version as equally authentic. Ready-made parallel corpora built from it include the Commission's DGT-TM translation memory (24 languages, latest release 2021), the JRC-Acquis corpus (22 languages) and MultiEURLEX (about 65,000 laws in 23 languages, with EuroVoc labels). To build a fresher or larger one, align EUR-Lex language versions on the [CELEX number](/glossary/celex/), then on article numbers, which are identical across languages. The text is reusable commercially under the EUR-Lex legal notice, with attribution.

## Why is EU law a parallel corpus at all?

[Regulation No 1/1958](https://european-union.europa.eu/principles-countries-history/languages_en) sets the languages of the EU institutions, and there are now 24 official and working languages. Regulations and directives are published in the [Official Journal](/glossary/official-journal/) in all of them. Irish was the last to reach full status: it became an official language in 2007 with a derogation, which [ceased to apply from 1 January 2022](https://eur-lex.europa.eu/eli/reg/2015/2264/oj).

The language versions are not translations of an authentic original. The Court of Justice said in *CILFIT* (Case 283/81, paragraph 18) that Community legislation is drafted in several languages and that the different language versions are all equally authentic, so interpreting a provision involves comparing them. For machine translation and multilingual models, that means the alignment is high quality by construction: every version was produced and checked by professional legal translators and lawyer-linguists to say the same thing.

The 24 languages are Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish and Swedish. Several of them, such as Maltese, Irish and Latvian, have little other high-quality parallel legal text available anywhere.

## Which ready-made parallel corpora exist?

| Corpus | Source | Languages | What it is | Terms |
|---|---|---|---|---|
| [DGT-TM](https://joint-research-centre.ec.europa.eu/language-technology-resources/dgt-translation-memory_en) | Commission, DG Translation; JRC | 24 | Sentence-aligned translation memory of the Acquis, public since November 2007; latest release 2021. Paper: Steinberger et al., LREC 2012 | Commission reuse policy (Decision 2011/833/EU), with attribution |
| [JRC-Acquis](https://joint-research-centre.ec.europa.eu/language-technology-resources/jrc-acquis_en) | JRC | 22 (v3.0) | Aligned EU legislation, about 1.06 billion words in total. Paper: Steinberger et al., LREC 2006 | Commission reuse policy (Decision 2011/833/EU) with attribution, per the JRC page; EuroVoc descriptors need the Publications Office's permission |
| [MultiEURLEX](https://huggingface.co/datasets/nlpaueb/multi_eurlex) | Chalkidis et al., EMNLP 2021 ([arXiv:2109.00904](https://arxiv.org/abs/2109.00904)) | 23 | About 65,000 EU laws with EuroVoc labels; document-level, built for classification | The card is internally inconsistent (CC BY 4.0 in its text, CC BY-SA 4.0 in its metadata) |
| [EUR-Lex-Sum](https://aclanthology.org/2022.emnlp-main.519/) | Aumiller et al., EMNLP 2022 | 24 | Acts paired with their summaries; only 375 acts are available in all 24 languages | See the dataset release |
| [Europarl](https://www.statmt.org/europarl/) | Koehn, MT Summit 2005 | 21 (v7) | European Parliament proceedings: debates, not legislation | See the release |
| [MultiLegalPile](https://arxiv.org/abs/2306.02069) | Niklaus et al., arXiv:2306.02069 | 24 | 689 GB of legal text from 17 jurisdictions; not aligned | Mixed licences by source; tagged CC BY-NC-SA 4.0 overall on Hugging Face |

The [OPUS](https://opus.nlpl.eu/) collection redistributes several of these (DGT, JRC-Acquis, EUbookshop, Europarl) in standard aligned formats. For multilingual legal classification, [LEXTREME (Niklaus et al., arXiv:2301.13126)](https://arxiv.org/abs/2301.13126) gathers 11 datasets in 24 languages.

## What are the limits of the ready-made corpora?

They are excellent baselines, and they have three limits for production use.

1. **They are frozen.** DGT-TM's latest release is 2021 and JRC-Acquis is older. Everything adopted since, including the AI Act, the Data Act and the Digital Services Act, is missing.
2. **They are sentence-level or document-level.** Translation memories lose the article structure; document-level datasets lose the alignment below the document. Neither keeps the identifiers a legal application needs to cite.
3. **Their terms vary.** The Commission corpora follow its reuse policy, but derived datasets add their own cards, some non-commercial. Check each one, and the underlying [EUR-Lex terms](/guides/eur-lex-data-licence-ai-training/), before commercial use.

## How do you build your own parallel corpus from EUR-Lex?

EU legal documents are aligned by design, which makes building your own corpus straightforward in principle.

1. **Enumerate documents by CELEX number.** The CELEX number (the GDPR is `32016R0679`) is the same for every language version of a document. [Cellar](/glossary/cellar/), the Publications Office repository behind [EUR-Lex](/glossary/eur-lex/), exposes the language manifestations of each work. See [CELEX, ECLI and ELI explained](/guides/celex-ecli-eli-identifiers-guide/).
2. **Fetch each language version** in a structured format (Formex XML or the HTML rendering) rather than PDF.
3. **Align at the article level first.** Article numbers, paragraph numbers and point letters are the same in every language, so a parser that splits each version into recitals, articles and annexes gives you aligned units without any statistical aligner.
4. **Align sentences within each article** with a standard sentence aligner. Because the units are short and already matched, errors stay local.
5. **Keep corrigenda in mind.** A corrigendum can correct one language version only. Record which version of each language you aligned, or a pair can become inconsistent after a correction.
6. **Use the same identifiers for consolidated texts.** Consolidated versions carry a CELEX number starting with 0 and ending in the consolidation date; align them with each other, not with the original act.

[EuroVoc](https://op.europa.eu/en/web/eu-vocabularies/eurovoc), the multilingual thesaurus managed by the Publications Office, is available in the 24 EU languages plus Albanian, Macedonian and Serbian, and gives you aligned subject labels as well.

## What about national law?

National legislation is monolingual, with a few exceptions such as Finland (Finnish and Swedish) and some bilingual regions. It is not a parallel corpus, but it is comparable data: a national measure that [transposes a directive](/guides/eu-directive-transposition-data/) covers the same subject as the directive, in the national language and the national drafting style. Transposition links give you directive-to-national pairs for cross-lingual retrieval and for training models that move between EU and national terminology. The sources and licences per country are in [National legislation datasets in Europe](/guides/national-legislation-datasets-europe/).

## How should you use parallel legal text in training?

| Use | How parallel EU law helps | Watch out for |
|---|---|---|
| Machine translation | High-quality, domain-specific sentence pairs for every EU language pair | Legal register only; mix with general data |
| Multilingual LLM pre-training | Professionally written text in low-resource EU languages | Heavy duplication of boilerplate across acts; deduplicate |
| Cross-lingual retrieval | Same-article pairs as positives across languages | Pair the same version and date in each language |
| Terminology | Aligned legal terms, and EuroVoc labels in 24 languages | Terms change between acts and over time |
| Evaluation | Held-out acts adopted after your training data | Older acts are likely already in web crawls |

## Key facts

- The EU has 24 official languages under Regulation No 1/1958; Irish has had full status since 1 January 2022.
- All language versions of EU legislation are equally authentic (CJEU, *CILFIT*, Case 283/81, paragraph 18).
- DGT-TM covers 24 languages (latest release 2021); JRC-Acquis v3.0 covers 22; MultiEURLEX has about 65,000 laws in 23 languages.
- The CELEX number is identical across language versions, and article numbering is identical too, so alignment can be done structurally.
- EUR-Lex content can be reused commercially under the EUR-Lex legal notice, with acknowledgement of the source.

## Frequently asked questions

### Is DGT-TM free for commercial use?

DGT-TM is published under the Commission's reuse policy, which permits commercial and non-commercial reuse free of charge with attribution. Check the conditions on the JRC page for the release you download.

### Are EU language versions translations of the English text?

Not legally. All language versions are equally authentic, and courts compare them when interpreting a provision. In practice most drafting happens in one language, mainly English, and the others are produced by translation and legal revision.

### How do I align EU documents across languages?

Use the CELEX number to match the document, then article, paragraph and point numbers to match the units, then a sentence aligner inside each unit. This avoids most of the errors of aligning whole documents statistically.

### Does any parallel corpus cover laws adopted after 2021?

The public Commission corpora stop at their release dates. For recent acts, build the alignment from EUR-Lex or use a maintained corpus that is updated as new acts are published.

## Where Arieta Data fits

Arieta Data's [corpus](/datasets/corpus/) stores one record per document and language, keyed by CELEX number for EU documents, with articles split and numbered so that language versions align at article level. In snapshot 2026-11 the [EU lane](/coverage/eu/) holds 61,410 documents, almost all of them in English: about 1,000 records (under 2 percent) are in other EU languages. The 15 national lanes add legislation in 14 national languages, and the [citation graph](/datasets/graph/) holds 66,670 transposition edges between directives and national measures. If you need specific EU language versions, tell us which, and we will tell you what the current snapshot holds and what we can add. National case law is included, delivered with personal data masked.

[Request a sample](/request-access/) in the languages you need. We reply within one business day, and a custom cut of the current snapshot by jurisdiction, language, date range or document type is ready within 5⁠–⁠6 business days.

*Not legal advice. Check each corpus's terms and the EUR-Lex legal notice before commercial use.*