Data Preparation Guide

Supported Datasets

EvoRM is validated on the ICEWS-WIKI dataset for entity alignment (EA) tasks.

ICEWS-WIKI (HHEA)

ICEWS-WIKI is a Highly Heterogeneous Knowledge Graph Alignment dataset where:

KG1 uses ICEWS relation IDs (e.g., 001, 002)
KG2 uses Wikidata relation IDs (e.g., P31, P17)
The relation schemas are completely different, making direct rule matching challenging

Data Structure

data/icews_wiki/
├── ent_ids_1          # KG1 entity ID → name mapping
├── ent_ids_2          # KG2 entity ID → name mapping
├── rel_ids_1          # KG1 relation ID → name mapping
├── rel_ids_2          # KG2 relation ID → name mapping
├── triples_1          # KG1 triples (h, r, t)
├── triples_2          # KG2 triples (h, r, t)
├── ref_pairs          # Reference (ground truth) entity pairs
├── ref_ent_ids        # Reference entity IDs
├── sup_pairs          # Supervision pairs (training)
├── message_pool/
│   ├── retriever_outputs.txt    # Candidate pairs (KG1→KG2, 5 candidates each)
│   ├── important_entities.txt   # Subset for LLM processing
│   ├── aligned_entities.txt     # Output aligned pairs
│   └── alignment_rules.txt      # Reference rules (optional)
└── evorm_state/                 # EvoRM persistent state
    ├── rules.json               # Distilled FOL rules
    └── hyperedges.json          # Hypergraph structure

Statistics

Metric	Value
Reference pairs	3,540
KG1 entities	3,540
KG2 entities	15,831
Candidate pairs	17,700 (5 per KG1 entity)
KG1 relations	256
KG2 relations	682

Adding New Datasets

To run EvoRM on a new dataset, ensure the following files are available:

ent_ids_1 / ent_ids_2: Tab-separated entity ID → name mapping
rel_ids_1 / rel_ids_2: Tab-separated relation ID → name mapping
triples_1 / triples_2: Tab-separated triples (head, relation, tail)
ref_pairs: Tab-separated ground truth pairs
message_pool/important_entities.txt: Candidate pairs for LLM processing

Then run:

from LLM1_label_selector import align_entities
aligned_pairs = align_entities(data_dir="./data/your_dataset")

DBP15K Datasets

The paper also evaluates on DBP15K variant datasets (EN-FR, EN-DE, ZH-EN, JA-EN, FR-EN). These datasets follow a similar structure and can be downloaded from:

DBP15K
JAPE