Datasets:

oeg
/

CQ2Onto

license: apache-2.0
language:
  - en
pretty_name: CQ2Onto
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - ontology-engineering
  - ontology-generation
  - term-extraction
  - benchmark
configs:
  - config_name: cq2onto
    data_files:
      - split: wine
        path: wine/cq_to_onto_wine.json
      - split: awo
        path: awo/cq_to_onto_awo.json
      - split: odrl
        path: odrl/cq_to_onto_odrl.json
      - split: water
        path: water/cq_to_onto_water.json
      - split: vgo
        path: vgo/cq_to_onto_vgo.json
      - split: swo
        path: swo/cq_to_onto_swo.json
  - config_name: cq2term
    data_files:
      - split: wine
        path: wine/cq_to_terms_wine.json
      - split: awo
        path: awo/cq_to_terms_awo.json
      - split: odrl
        path: odrl/cq_to_terms_odrl.json
      - split: water
        path: water/cq_to_terms_water.json
      - split: vgo
        path: vgo/cq_to_terms_vgo.json
      - split: swo
        path: swo/cq_to_terms_swo.json

CQ2Onto Benchmark & Dataset

Benchmark for evaluating LLM-assisted ontology generation from competency questions, across six domains. For each domain the dataset provides a gold OWL ontology, two CQ files (one per evaluation task), and the annotation spreadsheet used during construction. Each Ontology contains a set of CQs for CQ2Term, a set of CQs for CQ2Onto, a owl source file that representing all CQs for CQ2Onto, and an excel contains all annotation process. More details can be found here.

Two tasks:

CQ2Term: given a CQ, extract all possible classes and properties.
CQ2Onto: given a set of CQs, produce a full OWL ontology.

Dataset Construction

We have selected six ontologies in three diferent scales:

Ontology	Tier	Source CQs	Retained	New ⋆	CQ2Onto set	CQ2Term set
Wine	small	7	4	1	5	5
AWO	small	14	7	0	7	7
ODRL	medium	35	13	6	19	19
Water	medium	43	21	0	21	20
VGO	large	68	30	1	31	22
SWO	large	88	35	0	35	26

All sources of the selected ontologies:

File formats

Annotation Records:

<Domain>_CQs_Annotations.xlsx: annotation process with per-CQ class and property splits, plus axioms.

CQ2Onto Task:

cq_to_onto_<domain>.json (CQ2Onto Input): list of CQs. Gold standard is the ontology, corresponding to .owl file.

[
  {"id": "CQ1", "value": "Which wine characteristics should I consider when choosing a wine?"}
]

sub_<domain>.owl (CQ2Onto Gold Standard): OWL source code in RDF/XML. CQ-driven restriction of the source ontology, retaining only what's required to satisfy the CQs.

CQ2Term Task:

cq_to_terms_<domain>.json (CQ2Term Input & Gold Standard): list of CQs, with the gold standard class and property labels.

[
  {
    "id": "CQ1",
    "question": "Which wine characteristics should I consider when choosing a wine?", # Input Competency Question
    "classes": ["Wine", "WineDescriptor"], # Gold Standard Classes
    "properties": ["hasWineDescriptor"] # Gold Standard Properties
  }
]

Loading

from huggingface_hub import hf_hub_download
import json, rdflib

# CQ2Term: Load Dataset
path = hf_hub_download("oeg/CQ2Onto", "wine/cq_to_terms_wine.json", repo_type="dataset")
cqs = json.load(open(path))

# CQ2Onto
path = hf_hub_download("oeg/CQ2Onto", "wine/sub_wine.owl", repo_type="dataset")
g = rdflib.Graph().parse(path)

License

Apache 2.0.