OntoLAMA: LAnguage Model Analysis for Ontology Subsumption Inference

Dataset Summary

OntoLAMA is a set of language model (LM) probing datasets for ontology subsumption inference. The work follows the "LMs-as-KBs" literature but focuses on conceptualised knowledge extracted from formalised KBs such as the OWL ontologies. Specifically, the subsumption inference (SI) task is introduced and formulated in the Natural Language Inference (NLI) style, where the sub-concept and the super-concept involved in a subsumption axiom are verbalised and fitted into a template to form the premise and hypothesis, respectively. The sampled axioms are verified through ontology reasoning. The SI task is further divided into Atomic SI and Complex SI where the former involves only atomic named concepts and the latter involves both atomic and complex concepts. Real-world ontologies of different scales and domains are used for constructing OntoLAMA and in total there are four Atomic SI datasets and two Complex SI datasets.

Links

• Dataset Specification: https://krr-oxford.github.io/DeepOnto/ontolama/
• Zenodo Release: https://doi.org/10.5281/zenodo.6480540
• Paper: https://arxiv.org/abs/2302.06761 (Arxiv) or https://aclanthology.org/2023.findings-acl.213/ (ACL Anthology)

Languages

The text in the dataset is in English, as used in the source ontologies. The associated BCP-47 code is en.

Dataset Structure

Data Instances

An example in the Atomic SI dataset created from the Gene Ontology (GO) is as follows:

{
    'v_sub_concept': 'ctpase activity',
    'v_super_concept': 'ribonucleoside triphosphate phosphatase activity',
    'label': 1,
    'axiom': 'SubClassOf(<http://purl.obolibrary.org/obo/GO_0043273> <http://purl.obolibrary.org/obo/GO_0017111>)'
}

An example in the Complex SI dataset created from the Food Ontology (FoodOn) is as follows:

{
    'v_sub_concept': 'ham and cheese sandwich that derives from some lima bean (whole)',
    'v_super_concept': 'lima bean substance',
    'label': 0,
    'axiom': 'SubClassOf(ObjectIntersectionOf(<http://purl.obolibrary.org/obo/FOODON_03307824> ObjectSomeValuesFrom(<http://purl.obolibrary.org/obo/RO_0001000> <http://purl.obolibrary.org/obo/FOODON_03302053>)) <http://purl.obolibrary.org/obo/FOODON_00002776>)',
    'anchor_axiom': 'EquivalentClasses(<http://purl.obolibrary.org/obo/FOODON_00002776> ObjectIntersectionOf(<http://purl.obolibrary.org/obo/FOODON_00002000> ObjectSomeValuesFrom(<http://purl.obolibrary.org/obo/RO_0001000> <http://purl.obolibrary.org/obo/FOODON_03302053>)) )'
}

An example in the biMNLI dataset created from the MNLI dataset is as follows:

{
    'premise': 'At the turn of the 19th century Los Angeles and Salt Lake City were among the burgeoning metropolises of the new American West.',
    'hypothesis': 'Salt Lake City was booming in the early 19th century.',
    'label': 1
}

Data Fields

SI Data Fields

• v_sub_concept: verbalised sub-concept expression.
• v_super_concept: verbalised super-concept expression.
• label: a binary class label indicating whether two concepts really form a subsumption relationship (1 means yes).
• axiom: a string representation of the original subsumption axiom which is useful for tracing back to the ontology.
• anchor_axiom: (for complex SI only) a string representation of the anchor equivalence axiom used for sampling the axiom.

biMNLI Data Fields

• premise: inheritated from the MNLI dataset.
• hypothesis: inheritated from the MNLI dataset.
• label: a binary class label indicating contradiction (0) or entailment (1).

Data Splits

Source	#NamedConcepts	#EquivAxioms	#Dataset (Train/Dev/Test)
Schema.org	894	-	Atomic SI: 808/404/2,830
DOID	11,157	-	Atomic SI: 90,500/11,312/11,314
FoodOn	30,995	2,383	Atomic SI: 768,486/96,060/96,062 Complex SI: 3,754/1,850/13,080
GO	43,303	11,456	Atomic SI: 772,870/96,608/96,610 Complex SI: 72,318/9,040/9,040
MNLI	-	-	biMNLI: 235,622/26,180/12,906

Citation Information

The relevant paper has been accepted at Findings of ACL 2023.

@inproceedings{he2023language,
  title={Language Model Analysis for Ontology Subsumption Inference},
  author={He, Yuan and Chen, Jiaoyan and Jimenez-Ruiz, Ernesto and Dong, Hang and Horrocks, Ian},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2023},
  pages={3439--3453},
  year={2023}
}

Contact

Yuan He (yuan.he(at)cs.ox.ac.uk)