Chulinz/icj-citation-graph

ICJ Citation Graph

One graph holding every public decision of the International Court of Justice: 2,407 documents across 199 of the Court's 200 cases, from Corfu Channel (1947) to the decisions of 2026. Every document is parsed into structured text, split into 121,438 passages with precomputed embeddings, and linked to the case it belongs to, the earlier cases it cites, the treaty provisions it invokes, and the judges who sat on it.

Ships in two forms, use either:

Form	Files	Needs
Ready-made graph	neo4j.dump (1.7 GB)	Neo4j 5.26
Portable files	dataset/ (cases + provisions + documents + chunks JSONL + npy, 818 MB)	any language, numpy for vectors

Embeddings are included in both forms (natively on graph nodes in the dump, as embeddings.npy in the portable set), so nothing needs to be re-embedded.

Quick start

from datasets import load_dataset

# two configs -- documents and chunks have different schemas
docs = load_dataset("Chulinz/icj-citation-graph", "documents", split="train")
chunks = load_dataset("Chulinz/icj-citation-graph", "chunks", split="train")

Join them on chunks.document_id == documents.id (verified 1:N, no orphans on either side):

import pandas as pd
docs = pd.read_json("dataset/documents.jsonl", lines=True)
chunks = pd.read_json("dataset/chunks.jsonl", lines=True)
chunks["row"] = range(len(chunks))  # row i of embeddings.npy
merged = chunks.merge(docs, left_on="document_id", right_on="id", suffixes=("_chunk", "_doc"))

A runnable end-to-end walkthrough (load, inspect, join, reconstruct a document, semantic search) is in notebooks/dataset_demo.ipynb.

Or grab files directly:

hf download Chulinz/icj-citation-graph dataset/embeddings.npy --repo-type dataset --local-dir .

Portable files (dataset/)

cases.jsonl (199 rows)

The case-level lookup that documents.jsonl only references by case_id. Join documents.case_id -> cases.case_id to resolve a document's case name and span:

{"case_id": "070", "name": "Military and Paramilitary Activities in and against Nicaragua (Nicaragua v. United States of America)", "year_start": 1984, "year_end": 1991, "icj_documents": {"decision": 8, "pleadings": 8, "oral": 3, "other": 4}}

name is present on all 199 and year_start on all 199 (for pending cases, whose start the ICJ list page leaves blank, it is the first Court-act year scraped from the case page). A null year_end means the case is still pending -- the 22 rows with no year_end are exactly the 22 cases on the ICJ's official pending list (its 20 "Pending cases" plus the 2 under "Currently being heard or under deliberation"), so no separate status flag is needed.

icj_documents is a coverage snapshot of how many documents of each kind ICJ publishes for the case on its website (decision = judgments + orders + advisory opinions, pleadings = written proceedings, oral = verbatim records, other), scraped 2026-07-14 (see manifest.icj_documents_snapshot). It describes what exists upstream, not what is in this dataset: this dataset holds the decision documents (expanded into per-opinion rows), so the pleadings and oral counts show what a future harvest could still add. Corpus-wide that is ~1,850 pleadings and ~1,190 oral records not included here.

provisions.jsonl (357 rows)

The provision lookup that documents.jsonl only references by key in cited_provisions. Join cited_provisions -> provisions.key:

{"key": "statute:article_36", "title": "Article 36", "level": "article", "instrument": "statute", "text": "1. The jurisdiction of the Court comprises all cases which the parties refer to it ..."}

text is the full article body, present on 346 of 357. Empty on chapter containers and on a few citation-target keys that point at articles outside the four parsed instruments (e.g. statute:article_92, which is really a UN Charter article). The four instruments are the UN Charter, the ICJ Statute, the Rules of Court, and the Practice Directions.

documents.jsonl (2,407 rows)

One row per court document, citations folded in. A real row:

{
  "id": "ICJ_011_USNationalsMorocco_FRA_USA_1952-08-27_JUD_01_ME_00_EN",
  "case_id": "011",
  "title": "CASE CONCERNING RIGHTS OF NATIONALS OF THE UNITED STATES OF AMERICA IN MOROCCO (FRANCE v. UNITED STATES OF AMERICA)",
  "document_type": "judgment",
  "lang": "EN",
  "issued_date": "1952-08-27",
  "applicant": "FRANCE",
  "respondent": "UNITED STATES OF AMERICA",
  "stage": "merits",
  "filename": "ICJ_011_USNationalsMorocco_FRA_USA_1952-08-27_JUD_01_ME_00_EN.pdf",
  "source_url": "file:///app/.ocr_cache/pdf/ICJ_011_...pdf",
  "cited_cases": ["007", "008"],
  "cited_provisions": ["statute:article_38", "statute:article_40"]
}

Only id, case_id, title, document_type, lang, filename, cited_cases, cited_provisions are on every row. The rest appear when the document states them: issued_date on 1,429 rows, applicant/respondent on 574/572 (as raw header strings, often ALLCAPS -- the normalized country identities live in the graph's Country nodes, not here), stage on 635 (merits 344, preliminary_objections 266, compensation 2, plus a parsing artifact in on 23 rows).

document_type values: order (652), separate_opinion (507), dissenting_opinion (400), declaration (301), judgment (280), advisory_opinion (190), unknown (51), application (20), individual_opinion (6).

chunks.jsonl (121,438 rows)

One row per searchable passage:

{"id": "166-20240131-jud-01-00-en:0", "document_id": "166-20240131-jud-01-00-en", "seq": 0, "para": null, "zone": "title", "text": "..."}

embeddings.npy

float32 matrix of shape (121438, 1536), L2-normalized (cosine similarity = dot product). Row i corresponds to line i of chunks.jsonl. Model: gemini-embedding-001 with output_dimensionality 1536 (truncated from 3072 and re-normalized, per Google's guidance) -- embed queries the same way or similarities will be meaningless.

import json, numpy as np

emb = np.load("dataset/embeddings.npy", mmap_mode="r")
chunks = [json.loads(l) for l in open("dataset/chunks.jsonl")]
scores = emb @ query_vector  # query embedded with gemini-embedding-001 at 1536 dims, unit-normalized

manifest.json records the counts and the alignment contract.

Ready-made graph (neo4j.dump)

Full property graph with the same content plus the relationship structure. Vector indexes are part of the dump, so semantic search works immediately after restore, no re-indexing.

Host it with Docker

From the directory containing neo4j.dump:

# 1. load the dump into a fresh volume (~10 s)
docker run --rm -v icj_data:/data -v "$PWD":/backups:ro neo4j:5.26 \
  neo4j-admin database load neo4j --from-path=/backups --overwrite-destination=true

# 2. start the server on that volume
docker run -d --name icj-neo4j -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/choose-a-password \
  -e NEO4J_server_memory_pagecache_size=2G \
  -v icj_data:/data neo4j:5.26

# 3. verify (browser at http://localhost:7474, or:)
docker exec icj-neo4j cypher-shell -u neo4j -p choose-a-password \
  "MATCH (c:Case) RETURN count(c);"   # 199

Semantic search straight from cypher (index chunk_embedding, 1536-dim; embed the query with gemini-embedding-001 at 1536 dims, unit-normalized):

CALL db.index.vector.queryNodes('chunk_embedding', 10, $queryVector)
YIELD node, score
RETURN node.id, node.text, score

On an existing (non-Docker) Neo4j 5.26 install, stop the database and run the same neo4j-admin database load command.

Schema (node counts):

Node	Count	Key relationships
Case	199	Document -IN_CASE-> Case
Document	2,407	Document -CITES_CASE-> Case (6,070 edges)
Chunk	121,438	Chunk -PART_OF-> Document; embedding property, vector-indexed
Provision	357	Document -CITES-> Provision (7,939 edges); Provision -PART_OF-> Instrument
Instrument	4	UN Charter, ICJ Statute, Rules of Court, Practice Directions
Judge	220	Document -SAT_ON-> Judge
Country	112	Document -APPLICANT / -RESPONDENT-> Country

Citation links were extracted from the text of each decision and resolved to the correct case by name and year, so same-named cases from different eras never cross-link.

What the citation network shows

The ten most-cited cases in ICJ history, over the 6,070 case-to-case citation edges (1,301 documents cite at least one earlier case; 1,796 invoke at least one provision):

Case	Name	Times cited
070	Military and Paramilitary Activities (Nicaragua v. USA)	292
091	Application of the Genocide Convention (Bosnia v. Serbia)	206
001	Corfu Channel (UK v. Albania)	162
008	Interpretation of Peace Treaties	149
064	US Diplomatic and Consular Staff in Tehran	138
104	LaGrand (Germany v. USA)	128
048	Northern Cameroons	127
016	Anglo-Iranian Oil Co.	115
050	Barcelona Traction	111
095	Legality of the Threat or Use of Nuclear Weapons	110

Most-invoked provisions (of 7,939 document-to-provision citations): Statute art. 48 case management (643), art. 36 jurisdiction (471), art. 41 provisional measures (346), art. 40 institution of proceedings (259), Rules art. 79 preliminary objections (257).

What one case carries, using the most-cited one (Nicaragua v. USA, case 070) as the example: 32 documents of its own (judgments, orders, opinions, the application), 3,205 embedded passages, 22 judges linked across its phases, cites 21 earlier cases, and is cited by 292 later documents spread across 80 cases. Every case in the graph carries this same picture, queryable from either direction.

A fuller writeup with charts (documents by type, by decade) is in docs/eda.pdf.

Provenance and license

• Historical corpus (1947 - Oct 2023): the CD-ICJ corpus by Sean Fobbe (Zenodo, DOI 10.5281/zenodo.3826444), released under CC0 1.0. 2,285 documents.
• Recent decisions (Oct 2023 - 2026): scraped directly from icj-cij.org. 122 documents. ICJ decisions are public documents.
• The one absent case (case 2, from 1947) is unavailable at the source; its page on the ICJ website has returned a server error for years.

This dataset is released under CC0 1.0.

Known limitations

• 51 documents have document_type: "unknown" (old scans whose cover page was too degraded to classify); their text is fully searchable.
• Graph Country nodes are deduplicated and co-party lists split into individual states; four multi-word single entities (e.g. "Bosnia and Herzegovina") legitimately contain "and". The applicant/respondent strings inside documents.jsonl are the raw, un-normalized header text.
• 52 Judge nodes contain a comma: some are one judge plus a role suffix ("AGO, President of the Chamber"), some are several judges of a jointly-signed opinion parsed as one name. They do not affect the citation network.
• source_url is an internal provenance path (the corpus cache or scraper inbox), not a public link; the filename/id follows the CD-ICJ naming convention (case, parties, date, type, opinion number, language).
• issued_date is present on 1,429 documents; appended opinions usually do not print one.

How it was built

Each PDF is rendered page by page and read by an OCR model; a second pass classifies every line's role (title, paragraph, operative ruling, bench, signatures), yielding structured documents; a third pass extracts every citation of an earlier case and resolves it to a case id. Documents, passages, embeddings, and citation edges are written to Neo4j in one idempotent step.