WN18RR

Dataset Description

WordNet-based link prediction benchmark (improved) - derived from WN18 with test-set leakage removed

Original Source: https://figshare.com/ndownloader/files/21844185

Dataset Summary

This dataset contains RDF triples from WN18RR converted to HuggingFace dataset format for easy use in machine learning pipelines.

• Format: Originally tsv, converted to HuggingFace Dataset
• Size: 0.004 GB (extracted)
• Entities: 40,943
• Triples: 93,003
• Original License: Apache 2.0

Recommended Use

Lexical knowledge graph benchmarking, embedding evaluation

Notes\n\nStatic benchmark (2018). TSV format with pre-split train/valid/test files. Contains 11 relation types. Format: subjectrelationobject

Dataset Format: Lossless RDF Representation

This dataset uses a standard lossless format for representing RDF (Resource Description Framework) data in HuggingFace Datasets. All semantic information from the original RDF knowledge graph is preserved, enabling perfect round-trip conversion between RDF and HuggingFace formats.

Schema

Each RDF triple is represented as a row with 6 fields:

Field	Type	Description	Example
subject	string	Subject of the triple (URI or blank node)	"http://schema.org/Person"
predicate	string	Predicate URI	"http://www.w3.org/1999/02/22-rdf-syntax-ns#type"
object	string	Object of the triple	"John Doe" or "http://schema.org/Thing"
object_type	string	Type of object: "uri", "literal", or "blank_node"	"literal"
object_datatype	string	XSD datatype URI (for typed literals)	"http://www.w3.org/2001/XMLSchema#integer"
object_language	string	Language tag (for language-tagged literals)	"en"

Example: RDF Triple Representation

Original RDF (Turtle):

<http://example.org/John> <http://schema.org/name> "John Doe"@en .

HuggingFace Dataset Row:

{
  "subject": "http://example.org/John",
  "predicate": "http://schema.org/name",
  "object": "John Doe",
  "object_type": "literal",
  "object_datatype": None,
  "object_language": "en"
}

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("CleverThis/wn18rr")

# Access the data
data = dataset["data"]

# Iterate over triples
for row in data:
    subject = row["subject"]
    predicate = row["predicate"]
    obj = row["object"]
    obj_type = row["object_type"]

    print(f"Triple: ({subject}, {predicate}, {obj})")
    print(f"  Object type: {obj_type}")
    if row["object_language"]:
        print(f"  Language: {row['object_language']}")
    if row["object_datatype"]:
        print(f"  Datatype: {row['object_datatype']}")

Converting Back to RDF

The dataset can be converted back to any RDF format (Turtle, N-Triples, RDF/XML, etc.) with zero information loss:

from datasets import load_dataset
from rdflib import Graph, URIRef, Literal, BNode

def convert_to_rdf(dataset_name, output_file="output.ttl", split="data"):
    """Convert HuggingFace dataset back to RDF Turtle format."""
    # Load dataset
    dataset = load_dataset(dataset_name)

    # Create RDF graph
    graph = Graph()

    # Convert each row to RDF triple
    for row in dataset[split]:
        # Subject
        if row["subject"].startswith("_:"):
            subject = BNode(row["subject"][2:])
        else:
            subject = URIRef(row["subject"])

        # Predicate (always URI)
        predicate = URIRef(row["predicate"])

        # Object (depends on object_type)
        if row["object_type"] == "uri":
            obj = URIRef(row["object"])
        elif row["object_type"] == "blank_node":
            obj = BNode(row["object"][2:])
        elif row["object_type"] == "literal":
            if row["object_datatype"]:
                obj = Literal(row["object"], datatype=URIRef(row["object_datatype"]))
            elif row["object_language"]:
                obj = Literal(row["object"], lang=row["object_language"])
            else:
                obj = Literal(row["object"])

        graph.add((subject, predicate, obj))

    # Serialize to Turtle (or any RDF format)
    graph.serialize(output_file, format="turtle")
    print(f"Exported {len(graph)} triples to {output_file}")
    return graph

# Usage
graph = convert_to_rdf("CleverThis/wn18rr", "reconstructed.ttl")

Information Preservation Guarantee

This format preserves 100% of RDF information:

• ✅ URIs: Exact string representation preserved
• ✅ Literals: Full text content preserved
• ✅ Datatypes: XSD and custom datatypes preserved (e.g., xsd:integer, xsd:dateTime)
• ✅ Language Tags: BCP 47 language tags preserved (e.g., @en, @fr, @ja)
• ✅ Blank Nodes: Node structure preserved (identifiers may change but graph isomorphism maintained)

Round-trip guarantee: Original RDF → HuggingFace → Reconstructed RDF produces semantically identical graphs.

Querying the Dataset

You can filter and query the dataset like any HuggingFace dataset:

from datasets import load_dataset

dataset = load_dataset("CleverThis/wn18rr")

# Find all triples with English literals
english_literals = dataset["data"].filter(
    lambda x: x["object_type"] == "literal" and x["object_language"] == "en"
)
print(f"Found {len(english_literals)} English literals")

# Find all rdf:type statements
type_statements = dataset["data"].filter(
    lambda x: "rdf-syntax-ns#type" in x["predicate"]
)
print(f"Found {len(type_statements)} type statements")

# Convert to Pandas for analysis
import pandas as pd
df = dataset["data"].to_pandas()

# Analyze predicate distribution
print(df["predicate"].value_counts())

Dataset Format

The dataset contains all triples in a single data split, suitable for machine learning tasks such as:

• Knowledge graph completion
• Link prediction
• Entity embedding
• Relation extraction
• Graph neural networks

Format Specification

For complete technical documentation of the RDF-to-HuggingFace format, see:

📖 RDF to HuggingFace Format Specification

The specification includes:

• Detailed schema definition
• All RDF node type mappings
• Performance benchmarks
• Edge cases and limitations
• Complete code examples

Conversion Metadata

• Source Format: tsv
• Original Size: 0.004 GB
• Conversion Tool: CleverErnie RDF Pipeline
• Format Version: 1.0
• Conversion Date: 2025-11-06

Citation

If you use this dataset, please cite the original source:

Original Dataset: WN18RR URL: https://figshare.com/ndownloader/files/21844185 License: Apache 2.0

Dataset Preparation

This dataset was prepared using the CleverErnie GISM framework:

# Download original dataset
cleverernie download-dataset -d wn18rr

# Convert to HuggingFace format
python scripts/convert_rdf_to_hf_dataset.py \
    datasets/wn18rr/[file] \
    hf_datasets/wn18rr \
    --format tsv

# Upload to HuggingFace Hub
python scripts/upload_all_datasets.py --dataset wn18rr

Additional Information

Original Source

https://figshare.com/ndownloader/files/21844185

Conversion Details

• Converted using: CleverErnie GISM
• Conversion script: scripts/convert_rdf_to_hf_dataset.py
• Dataset format: Single 'data' split with all triples

Maintenance

This dataset is maintained by the CleverThis organization.