# Citation graphs as supervision signal for legal retrieval

> A legal citation graph links documents by explicit, typed relations such as amendment, repeal, legal basis and transposition, so its edges can serve as positive pairs, hard-negative sources and gold relevance sets for legal retrievers without manual labelling; Arieta Data's graph dataset types every edge and records where it came from, with no probabilistic edges.

Source: https://data.arietalabs.com/guides/citation-graphs-legal-retrieval-supervision/
Published: 2026-11-13 · Updated: 2026-11-14 · Author: Nejc Fošnarič, Founder, Arieta Labs (https://data.arietalabs.com/about/#nejc-fosnaric/)

A legal citation graph is useful supervision for retrieval because legal documents state their relations explicitly. An EU act names the acts it amends, repeals and is based on; a judgment names the provisions and judgments it relies on; a member state notifies the measures it adopted to transpose a directive. Each of these relations is a labelled pair of documents that no annotator had to produce. Used carefully, the edges of a [citation graph](/glossary/citation-graph/) give positive pairs for contrastive training, structured sources of hard negatives, and gold sets for retrieval evals. The care lies in edge types, edge provenance and time: a citation is not always a relevance judgment, and a graph is never complete.

## What is a legal citation graph?

A legal citation graph is a directed graph whose nodes are legal documents and whose edges record that one document cites, amends, repeals, implements or transposes another. In the EU, nodes are identified by [CELEX numbers](/glossary/celex/); national documents carry their own ids and, where available, an [ELI](/glossary/eli/) or [ECLI](/glossary/ecli/).

Much of the EU graph is published by the Publications Office. [Cellar](/glossary/cellar/), the repository behind [EUR-Lex](/glossary/eur-lex/), records structural relations between acts (amends, repeals, corrects), citations between [CJEU](/glossary/cjeu/) judgments, and the national measures member states notify for directives. The rest comes from the "notice" pages of EUR-Lex documents, which list legal bases, proposals and the delegated and implementing acts that complete an act.

## Why do citations work as supervision?

Citations work as supervision because they are document-level relevance signals created by experts for their own purposes. The same idea is established outside law: [SPECTER (Cohan et al., ACL 2020)](https://aclanthology.org/2020.acl-main.207/) trained scientific-document embeddings with a citation-informed triplet loss, using cited papers as positives and uncited papers as negatives.

Legal retrieval research has used the same signal, mostly on US and Canadian case law:

| Work | Data | Supervision from citations |
|---|---|---|
| [Huang et al., ICAIL 2021](https://arxiv.org/abs/2106.10776) | US Board of Veterans' Appeals decisions | Recommend the authority to cite at a point in a draft decision, from its context |
| [LePaRD (Mahari et al., ACL 2024)](https://aclanthology.org/2024.acl-long.532/) | US federal judicial citations to precedent | Predict the cited precedent passage from the citing context |
| [CLERC (Hou et al., Findings of NAACL 2025)](https://aclanthology.org/2025.findings-naacl.441/) | US federal case law from the Caselaw Access Project | Retrieve the cases cited by a passage of legal analysis |
| [COLIEE](https://coliee.org/) | Canadian case law and Japanese civil code | Task 1 asks systems to retrieve the cases "noticed" by a query case |

EU legislation has a property case-law citations lack: many of its edges are typed. A retrieval model that knows that act A amends act B, rather than merely mentions it, can learn different notions of relevance for different query intents.

## Which edge types matter, and what do they teach?

Each edge type supports a different retrieval task. Arieta Data's [graph dataset](/datasets/graph/) uses eight types. In snapshot 2026-11 it holds 801,149 edges: 651,075 `reference`, 66,670 `transposition`, 37,922 `amendment`, 33,900 `legal-basis`, 9,364 `implementation`, 10 `corrigendum`, 1,444 `repeal` and 764 `proposal`.

| Edge type | Meaning | Supervision use |
|---|---|---|
| `reference` | The citing document refers to the target | General relevance positives; context-to-target citation prediction |
| `amendment` | The citing act amends the target, including partial replacement | "What changed this act?" retrieval; pairs with version history |
| `repeal` | The citing act repeals the whole target act | Filtering repealed law; negatives for "current law" queries |
| `corrigendum` | The citing document corrects a language version of the target | Deduplication; language-version alignment |
| `implementation` | A delegated or implementing act that completes the target | Retrieving the detailed rules under a framework act |
| `legal-basis` | The target is the legal basis of the citing act | Treaty-article to secondary-law retrieval |
| `proposal` | A Commission proposal and the act it would amend or became | Linking legislative history to final text |
| `transposition` | A national measure notified for, or recorded as transposing, a directive | Cross-jurisdiction and cross-lingual retrieval from directive to national law |

Two rules from the source data change how edges should be read. A `repeal` means the whole act: a partial repeal or replacement of an article or annex is typed `amendment`. And where one act both amends and corrects another, the stored edge carries the stronger type, so each pair appears once.

## Why does edge provenance matter?

Edge provenance tells you how much to trust an edge and what population it represents. Arieta Data records an `origin` on every edge:

| Origin | Where the edge comes from |
|---|---|
| `notice` | The EUR-Lex notice page of the citing document |
| `cellar` | A relation recorded in Cellar, including case-law citations |
| `notice+cellar` | Both of the above |
| `cellar-nim` | A national measure a member state notified to the Commission for an EU act |
| `source` | A national source that records the relation and names the target by an id |
| `resolver` | A textual reference in a national source matched by a deterministic resolver |

There are no probabilistic edges. An edge exists only when the source records it or a deterministic, unique match was made, so you do not inherit another model's errors as labels. The `cellar-nim` edges deserve one caution: a [transposition](/glossary/transposition/) notification is a member state's statement that a measure relates to a directive, not a finding that it transposes the directive completely or correctly. How transposition data is structured is set out in [EU directive transposition as structured data](/guides/eu-directive-transposition-data/).

## How do you turn the graph into training data?

A citation graph supports at least four training and evaluation patterns for legal retrievers.

**Positive pairs.** Use the citation `context` (the text around a citation in the citing document, where captured) as the query and the cited document or article as the positive. For legislation, pair an article of an implementing act with the framework act it completes.

**Hard negatives.** Good negatives are close but wrong. The graph gives several principled sources: acts that share a legal basis with the positive but are not cited; earlier acts repealed by the positive; siblings amended by the same amending act; and, with version history, the pre-amendment version of the cited article. Remove any candidate that is itself connected to the query document, so that a true positive is not mislabelled as a negative.

**Graph-aware reranking.** Features such as "is a candidate amended by a later document in the result set" or "is this candidate repealed" let a reranker demote outdated law. The repeal and amendment edges do this work deterministically.

**Gold sets for evaluation.** For a directive, the set of notified national measures is a ready-made gold set for directive-to-national retrieval across languages. For a framework regulation, its implementing and delegated acts form another. These are the kinds of gold sets [Arieta Data eval sets](/datasets/eval/) are built from, to your spec on request. The full method is in [How to build a legal eval set from a citation graph](/guides/legal-eval-set-citation-graph/).

## What are the pitfalls?

Citation supervision fails in predictable ways, and each has a mitigation.

1. **Incompleteness.** No graph holds every relation. Edges from EUR-Lex notice pages cover the documents that were extracted, not the whole of EUR-Lex. Treat an absent edge as unknown, not as a negative, unless you have checked it.
2. **Citation is not relevance.** A boilerplate reference to a treaty article is not evidence that the article answers a query. Filter by edge type and weight `reference` edges lower than structural ones.
3. **Temporal leakage.** An amendment edge from 2026 must not be visible to an eval item dated 2025. Join edges with the [history dataset](/datasets/history/) and the method in [Point-in-time legal data](/guides/point-in-time-legal-data-version-history/).
4. **Popularity bias.** Heavily cited acts dominate positives. Sample by edge, by document and by type to balance the training set.

## Key facts

- Legal citations are explicit and typed: an act names what it amends, repeals and is based on, and member states notify their transposition measures.
- Citation-informed training is established outside law (SPECTER, ACL 2020) and in legal retrieval research (LePaRD, CLERC, COLIEE).
- Arieta Data snapshot 2026-11: 801,149 typed edges across EU and national law, each with a recorded origin and no probabilistic edges.
- An absent edge is unknown, not a negative; a citation is not always a relevance judgment.
- Join edges with version history to avoid temporal leakage.

## Frequently asked questions

### What is a hard negative in legal retrieval?

A document that looks relevant but does not answer the query, such as an act with the same legal basis that the query document does not cite, or the pre-amendment version of the cited article. The graph supplies these in a principled way.

### Can I train a retriever on citations alone?

Citations give strong positives, but they are incomplete and some are boilerplate. Filter by edge type, down-weight generic references and combine graph supervision with a reviewed evaluation set.

### Does the graph include national law?

Yes. Arieta Data's graph links EU legislation, CJEU case law and national legislation, including edges recorded by national sources and national transposition measures. National case law is included, delivered with personal data masked.

## Where Arieta Data fits

Arieta Data's graph dataset is a typed, directed edge list linking EU legislation, CJEU case law and national legislation, with `source`, `target`, both jurisdictions, `type`, `origin`, citation `context` where captured and the `snapshotId`. It is cut from the same monthly [snapshot](/glossary/snapshot/) as the [corpus](/datasets/corpus/) and the history, so edges, text and versions join on the same ids. National case law is included, delivered with personal data masked. Snapshot 2026-11 has 801,149 edges; per-lane counts are on the [coverage page](/coverage/), with the [licensing page](/licensing/) for terms, and [request access](/request-access/) for the sample: we reply within one business day, and a custom cut is ready within 5⁠–⁠6 business days. For what else exists in the market, read [EU legislation datasets for LLM training](/guides/eu-legislation-datasets-llm-training/); for using retrieved articles in RAG, read [How to chunk legal documents for RAG](/guides/legal-rag-chunking-article-level/).

*Not legal advice. A citation or transposition edge records a relation stated by a source; it is not a legal assessment.*