Not all schemas are the same. Some use weak semantics, which are broad categories with implied meanings. Others use strong semantics, which are explicit hierarchies with formal relationships. The difference is important.

Research from SciBite shows this distinction through an experiment in extracting life sciences articles from Wikipedia. Wikipedia organizes articles using categories, a form of weak semantics similar to SKOS (Simple Knowledge Organization System). Categories point upwards and downwards to indicate more general or more narrow groupings. But the specific meaning is implicit.

 This works for people browsing the site, but it causes problems for machine learning. When researchers tried to pull all the life sciences articles by following Wikipedia’s category tree, they got some strange results. The “Hearing” category led to “Sound,” which led to “Music by country,” and finally to Peruvian folk music. The “Cocaine” category led to “Fictional cocaine users,” and then to Sherlock Holmes. Neither of these belongs in a life sciences dataset.

‍

‍

Strong semantics solve this problem. In an ontology, the relationships between classes are explicit. The World Wide Web Consortium (W3C) has established the Web Ontology Language (OWL) as a standard for this. What is true for a given class is also true for all its subclasses. This allows for better reasoning and validation. If a dataset is supposed to cover all central nervous system diseases but is missing examples of Alzheimer’s, the ontology can flag that gap.

For multi-domain AI, strong semantics are essential. They provide a solid foundation for checking dataset completeness and aligning schemas across different domains.

Ontologies do more than just organize concepts. They can also be used to validate datasets. A research paper published on arXiv suggests using two types of ontologies to make sure training datasets for safety-critical applications like autonomous driving are complete and robust.

Domain Ontologies

A domain ontology defines all the important concepts in a specific area. For emergency vehicle detection, the ontology might include ambulances, fire trucks, and police cars. Each concept is clearly defined, and the relationships between them are formalized.

The ontology acts as a checklist. Does the dataset have examples of all the vehicle types? Are all the subtypes represented? If the ontology includes “ambulance with lights on” and “ambulance with lights off” as separate concepts, the dataset needs to have both.

This approach can be used in any domain. A medical imaging ontology could define anatomical structures and pathologies. An autonomous driving ontology could define road users and infrastructure. Each ontology provides a formal way to specify what the dataset needs to cover.

Image Quality Ontologies

Having all the right concepts is not enough. A dataset might have all the right objects but not capture the variations in quality that a model will see in the real world. An image quality ontology can help with this.

An image quality ontology defines different quality dimensions, like lighting conditions (day, night), weather (clear, rain, snow), and occlusions (partial, full). For each concept in the domain ontology, the dataset should have examples across all these quality dimensions.

A model trained only on daytime images will not work well at night. A model trained only in clear weather will fail in the fog. By formalizing these quality dimensions, we can systematically check for gaps in the dataset.

The two-ontology approach increases trust in ML models used in safety-critical domains. Because ML models embody what they are trained with, ensuring the completeness of training datasets increases trust in the training of ML models.