💬 NLP

Automatic Construction of a Legal Citation Graph from 100 Million Ukrainian Court Decisions: Large-Scale Extraction, Topological Analysis, and Ontology-Driven Clustering

This paper presents the automatic construction of a massive legal citation graph from over 100 million Ukrainian court decisions, revealing that unsupervised analysis of 502 million citation edges effectively encodes legal domain boundaries, predicts legislative importance with near-perfect accuracy, and detects regime changes, thereby enabling the creation of an ontology-driven workflow for LLM-assisted legal analysis.

Original authors: Volodymyr Ovcharov

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Volodymyr Ovcharov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the entire legal system of Ukraine as a massive, bustling library containing over 100 million books (court decisions). For years, these books sat on shelves, mostly unread by computers, because the way they referenced other laws was messy, inconsistent, and written in a complex human language.

This paper describes a project where the authors built a super-fast robot librarian to read every single one of those books and map out exactly how they connect to one another.

Here is the breakdown of what they did and what they found, using simple analogies:

1. The Robot Librarian (The Extraction)

The authors created a specialized tool (a "regex pipeline") that acts like a high-speed scanner.

The Task: It had to find references to laws inside 100.7 million court decisions. That's 1.1 terabytes of text—enough to fill a small library.
The Speed: It did this in just 5 hours on a standard computer server. To put that in perspective, a human reading one page a minute would take centuries to do this.
The Result: The robot found 502 million connections (citations). It was incredibly accurate, getting a perfect score on a test sample. It successfully identified six different types of references, from specific articles in the Civil Code to rulings by the Supreme Court.

2. The Map of Connections (The Graph)

Once the robot finished reading, the authors didn't just have a pile of notes; they built a giant spiderweb (a graph).

How it works: Imagine every court decision is a dot, and every law it mentions is another dot. The robot drew a line between them.
The Shape: They found that this web follows a "Power Law." In plain English, this means a tiny handful of laws are super-popular (cited millions of times), while most laws are cited very rarely.
- Analogy: Think of it like a social media network. A few celebrities (like the main articles on theft or civil procedure) have millions of followers, while most regular people have only a few. In Ukraine, the "celebrity" laws are the procedural rules that judges use in almost every single case.

3. Discovering Hidden Neighborhoods (Ontology & Clustering)

The most surprising part of the paper is what happened when they looked at how laws are cited together.

The Discovery: They used a computer algorithm to group laws that are frequently cited in the same cases. They didn't tell the computer what "Civil Law" or "Criminal Law" was. They just let the data speak.
The Result: The computer naturally grouped the laws into four distinct neighborhoods: Civil, Criminal, Administrative, and Commercial.
- Analogy: Imagine walking into a room full of people and asking them to sort themselves into groups without telling them the rules. If you ask them to group by "who they talk to," the accountants will naturally cluster together, and the artists will cluster together. The computer did this with laws, proving that the structure of the court system itself reveals the categories of law, even without human experts labeling them.

4. Reading the History (Time Travel)

Because they had data from 2007 to 2026, they could watch the legal system change in real-time.

Reforms: When a new law code was introduced (like in 2012 or 2017), the map showed a sudden "earthquake." Old laws stopped being cited, and new ones exploded in popularity.
The 2022 Invasion: The paper notes a specific "spike" in 2022. The variety of laws being cited suddenly became much more chaotic and diverse (entropy increased). New laws about the war appeared on the map for the first time, showing how the legal system instantly adapted to the crisis.

5. Predicting the Future

The authors tested if they could guess which laws would become important in the future.

The Test: They used the history of citations to predict the top 1,000 most important laws for the years 2020–2026.
The Score: They were 99.8% accurate.
The Lesson: The best way to know if a law is important isn't to read the text; it's to see how often judges cite it. The "popularity" of a law is a perfect predictor of its future importance.

Why This Matters (According to the Paper)

The authors explain that this map is now being used to help Artificial Intelligence (AI) understand Ukrainian law.

Currently, AI models often hallucinate or make things up because they don't have a solid "map" of how laws relate to each other.
This project provides that map. It gives the AI a "domain layer"—a structured, data-driven guide that tells the AI, "These laws belong together, and this is how they are used in real life." This helps the AI give more accurate, grounded legal advice.

In summary: The paper is about building a massive, automated map of Ukraine's legal system. By letting the data draw the lines between laws, they discovered that the court system naturally organizes itself into clear categories, that a few laws hold the whole system together, and that this map can teach AI how to think like a lawyer.

Technical Summary: Automatic Construction of a Legal Citation Graph from 100 Million Ukrainian Court Decisions

Problem Statement
The Unified State Register of Court Decisions (EDRSR) represents the largest open judicial corpus in continental Europe, containing over 100 million decisions. Despite its size, this resource has remained largely unexploited for computational legal analysis. Existing legal citation network research has focused on common-law jurisdictions (e.g., U.S., Dutch, Danish) where citation conventions are explicit and datasets are smaller (10³–10⁵ decisions). Continental legal systems like Ukraine's present distinct challenges: citations target legislation articles rather than prior cases, formats are inconsistent due to morphological variations and abbreviations, and the sheer volume (1.1 TB of text) requires industrial-scale processing. No prior work has attempted citation extraction at the 100-million-decision scale.

Methodology
The paper proposes a three-stage pipeline: large-scale extraction, topological analysis, and ontology-driven clustering.

Large-Scale Extraction Pipeline:
- Data Source: 100.7 million court decisions (99.5 million full texts) from the EDRSR, spanning 2000–2026.
- Technique: A regex-based pipeline processes the data partition-by-partition using Python multiprocessing and server-side PostgreSQL cursors.
- Target Types: Six citation types are extracted: (1) Codex article references (e.g., "ст. 625 ЦК України"), (2) Named law references, (3) Constitutional references, (4) Inter-case references, (5) Law-by-number references, and (6) Supreme Court ruling references.
- Performance: The pipeline processes approximately 200,000 rows/second on a single 16-core server, completing the extraction of 502 million citation links in roughly 5 hours.
- Validation: On a manually annotated sample of 200 decisions (1,903 citations), the pipeline achieved 100% precision across all six types. A recall proxy of 0.86 was observed when cross-checking against stored records, with the 14% gap attributed to normalization differences (e.g., article range expansion) rather than missed substantive citations.
Graph Construction:
- Bipartite Graph ( $G_B$ ): Nodes represent decisions ( $D$ ) and legislation articles ( $L$ ). Edges connect a decision to the legislation it cites.
- Co-citation Projection ( $G_L$ ): A unipartite graph where legislation articles are connected if cited together in the same decision. Edge weights reflect the frequency of co-citation, serving as a proxy for semantic relatedness.
- Decision Similarity Graph ( $G_D$ ): A graph connecting decisions that cite common legislation (computed lazily).
Analysis and Ontology Construction:
- Community Detection: The Louvain algorithm is applied to the co-citation projection to identify clusters of legislation articles.
- Ontology Derivation: These clusters form an automatically constructed legal ontology where classes correspond to legal domains (e.g., Civil, Criminal) and individuals are the specific articles. This approach relies on the distributional hypothesis ("you shall know a law by the decisions that cite it") without requiring labeled data or embedding models.

Key Results

Scale and Efficiency: The study successfully extracted 502 million citation edges connecting decisions to 18.4 million unique legislation articles. This dataset is three to four orders of magnitude larger than previous legal citation studies.
Topological Properties:
- Power Law Distribution: The degree distribution of legislation articles follows a power law with exponent $\alpha = 1.57 \pm 0.008$ . This places the Ukrainian network between the EU Court of Justice ( $\alpha \approx 1.7$ ) and the U.S. Supreme Court ( $\alpha \approx 2.1$ ), indicating a heavy tail where a small set of "hub" articles (e.g., procedural codes) are cited by millions of decisions.
- Centrality Divergence: PageRank centrality diverges significantly from raw citation counts (Spearman $\rho = 0.70$ ). For instance, Article 19 of the Constitution ranks 42nd by raw count but 3rd by PageRank, reflecting its structural role as a bridge between legal domains.
Ontological Recovery: Community detection on the co-citation graph successfully recovers established legal domains (Civil, Criminal, Administrative, Commercial) with modularity scores ( $Q$ ) between 0.44 and 0.55. The structure shows high temporal stability (Normalized Mutual Information $NMI = 0.83–0.86$ across four-year periods), confirming the graph as a durable representation of the legal order.
Predictive Power: A logistic regression model using citation features (historical volume, growth ratio, etc.) predicts the top-1,000 most cited articles in the future with an AUC of 0.9984, significantly outperforming a naive frequency baseline.
Temporal Dynamics and Crisis Detection:
- Legislative Reforms: The graph captures quantitative phase shifts corresponding to major reforms, such as the 2012 Criminal Procedure Code launch and the 2017 judicial reform.
- War Impact: The 2022 Russian invasion is detected as a citation entropy spike ( $H: 11.02 \to 13.49$ ) and the emergence of new nodes related to wartime legislation (e.g., Criminal Code Art. 111-1 regarding collaboration).

Significance and Claims
The paper claims three primary contributions:

Scalable Extraction: It demonstrates that industrial-scale legal NLP can be achieved on commodity hardware using simple regex-based methods, achieving high precision without specialized infrastructure.
Data-Driven Ontology: It provides the first machine-readable ontology for Ukrainian statute relationships, derived entirely from judicial practice rather than expert curation. This addresses a structural gap in ontology-controlled LLM systems, providing a domain layer that grounds legal reasoning in actual usage patterns.
Quantitative Legal History: The citation graph serves as a high-resolution record of legislative regime transitions, capable of detecting systemic changes (reforms, invasions) through citation entropy and density shifts.

The authors position this work as an extension of the knowledge extraction and distributional semantic modeling programs proposed by Palagin et al., applied here to a scale and domain previously unattempted in Ukrainian NLP. The resulting ontology is operationalized as the domain layer for workflow memory systems in LLM-assisted legal analysis, enabling informed human corrections of AI outputs. The extraction pipeline, analysis code, and aggregated statistics are released as open data.