OASF Ontology and Semantics

[MSand53]

A space for exploring how formal knowledge representation, ontologies, and semantic technologies offer real value with OASF

[MSand53]

From what I can gather from the OASF site pages – Skills, Domains and Modules – there is a controlled vocabulary structured with a hierarchy to classify and retrieve parent-child relationships. This forms a Taxonomy with consistency and standardisation that makes predictable browsing and reduces ambiguity. NLP is the largest group in the catalogue with 44 items in the Taxonomy. For most cases and most users this Taxonomy is sufficient for most use cases. The purpose of this Slack Group is to consider whether a further and more formal model, such as an ontology, would offer any value or support a specific user group.
Taxonomies are closed by nature, returning only the detail in their structure. An ontology, with its formal semantics-based knowledge representation, may offer more to Agents, especially in a decision support model or a specific chain to reach an outcome. A Taxonomy provides the Is-A relationship, whereas an ontology offers multiple inheritance and complex classifications such as Has-A, Part-Of, Inverse, Disjoint, and Transitive. It also provides constraints through Property assertions. For example:
Currently:-
Automatic Speech Recognition Is-A Audio Processing Is-A Multi-modal
Property Name: ID, Type = Integer
Property Name: Name, Type = String

Potentially through an ontology in Turtle format:-
oasf: AutomaticSpeechRecognition rdf:type oasf:Skill ;
   oasf:hasInputModality "audio" ;
   oasf:hasOutputModality "text" ;
   oasf:hasLatencyClass "real-time" ;
   oasf:maxLatencyMs 100 .

This can be extended in order to qualify how a Skill is defined eg.

Modality axiom (what kind of data)
·      Text, Image, Audio, Video, Tabular, Multimodal
Functional axiom (what kind of operation)
·      Understanding, Generation, Classification, Transformation, Retrieval, Reasoning
Non-functional axiom (operational constraints)
·      Latency class (real-time, batch)
·      Privacy risk (none, low, high)
·      Resource requirements (CPU-only, GPU-required)
·      Accuracy class (approximate, high-precision)
·      Cost class (free, metered, expensive)

[MSand53]

I am working on agent-based security solutions with EV and EVSE technologies so I have used my current example to illustrate example requirements:

Agent solution requirement:

• An engineer photographs an installed charging station
• Agent System must identify installation errors in real-time (while they're on site)
• Must explain the error (generate text)
• Privacy-sensitive (can't send images to cloud)
• Interaction through a handheld mobile device

we are looking at a SPARQL query to identify agent via:

SELECT ?agent WHERE {
    ?agent oasf:hasSkill ?skill .
    ?skill oasf:hasInputModality "image" ;
           oasf:hasOutputModality "text" ;
           oasf:hasLatencyClass "real-time" ;
           oasf:privacyRisk "none" ;  # must be local/edge processing
           oasf:hasFunction "classification" .
}

This returns an Agent capable of image classification + text explanation. This could be extended with speech interaction and advisory capability with appropriate knowledgebase