An Ontology Design Pattern to Define Explanations | Tiddi et al. 2020, K-CAP

Page numbering is not the original from the paper I think; I couldn't get access and read a pre-print version.

Background of this paper: the authors (in their overall work) aim "to automatically identify explanations from data agnostically from the application domain" which first requires "a formal definition of what an explanation is" (1). The explanandum for them is "data patterns". The goal is to leverage machine-accessible knowledge of the "Web of Data" instead of relying on human experts and domain knowledge; hence, "agnostically".

So the authors survey various definitions of explanation in the cognitive sciences, broadly construed.

They then design an "ontology design pattern" to model these explanations and, specifically, use this ontology to instantiate the concepts of explanation in the analysed disciplines.

So if I understand correctly, this work models the various aspects that appear in the concept of explanation. By providing instantiations, we can clearly see commonalities and differences with other instantiated concepts.

This ontology is then usable to "analyse the validity of the explanations produced by our, and other, frameworks." (1)

The most obvious limitation of deriving explanations for data patterns is, since this is a data-driven approach, how to ensure the explanation is causal instead of explaining by (spurious) correlations.

"In a broader sense, in this work we aim at identifying which are the important actors of an explanation, i.e. which elements and interactions are needed to let us declare that, given a pattern and the derived observation, we have raised them to an explanatory level, not only to the one of a pure correlation." (3)

I find this relatively optimistic. How is having a formal description (in terms of an ontology) useful for assessing whether an (data-derived?) explanation is causal?

Survey

Looks interested, skipped for now due to time constraints.

The summary of different notions of explanation in different fields is nice though, so I'll quote it here for reference:

Philosophy is the "discipline of asking questions and checking answers" [16]. An explanation in Philosophy is a process where a set of initial elements (an event and some initial conditions) are described, deduced or put into a relation with an output phenomenon according to a set of (empirical or metaphysical) laws. (5)

Psychology tries to understand the human (and animal) cognitive processes. Here, E is intended as a model of components (entities and activities) that mechanically interact to produce regular changes from some starting conditions to a termination condition. (5)

Neuroscience gives an understanding of the brain. Explanations are seen as generalisations that describe relation- ships between input and output variables, that are stable under some appropriate range of interventions without any law nor sufficient condition. (5)

Computer Science deals with computer programs as a medium to perform human operations. Explanations are generalisations of some starting knowledge (some facts and some prior knowledge) mapped to another, new knowledge under constraints of certain criteria. (5)

Sociology focuses on the links between the human’s cognitive processes and his sociocultural world. It intends explanations as the act of giving a meaning to some social regularities occurring because of some social behaviours. (5-6).

Linguistics provides theories dealing with the nature of the human language. An explanation is the process of deriving the regularities of a language (the grammar) from a set of linguistic-specific facts respecting some wellformedness constraints. (6)

The general abstracted formalisation for the concept of explanation is: "When X happens, then, due to a given set of circumstances C, Y will occur because of a given law L." (6)

Explanation design pattern

Finally, near the end of the paper an explanation of what an Ontology Design Pattern is (ODP) :-).

Ontology Design Patterns are small, motivated ontologies to be exploited as building blocks in ontology design [15]. Their purpose is to be used as designing components of a problem with the assumption that an ontology should be modelled according to those patterns, with appropriate dependencies between them, plus some necessary design expansion based on the specific situation’s needs. We followed this idea to build an ontology design pattern that represents an explanation. We reused some of the existing design patterns and ontologies, and then extended classes and properties according to our needs. (6)

I'm not sure I get it. So the Ontology Design Pattern is itself an ontology, that is used to describe how other ontologies should be construed?

 Tiddi (2015) - explanation ontology design pattern

  • Main class Explanation:
    • This is the class that will be instantiated
    • Constraints specified as OWL class restrictions
  • Those constraints:
    • Explanation is of antecedent event expressed as hasExplanans relation with part:Event (comes from Participation ODP; can represent phenomena, facts, variables, observations, knowledge...)
      • Antecedent: before; explanans is the thing that does the explaining.
    • posterior event: hasExplanandum with part:Event.
    • a context that relates the event: hasCondition relation with situation:Situation
      • the idea is that explanans and explanandum should share the same context to avoid spurious associations
      • it is assumed we can infer for a part:Event whether it fits a situation:Situation through situation:hasSetting; this is expressed through a "SubPropertyChain" in "Manchester Syntax".
  • dul:Theory: see DOLCE+DnS Ultralite Ontology.
  • dul:Agent who produces the explanation (also see DOLCE+DnS Ultralite Ontology)?
    • Explanation relates with dul:Agent through dul:isConceptualizedBy.

Pattern application to frameworks that produce explanations

The authors use a framework called Dedalo:

Dedalo wants to explain the fact P that a group of users is searching the Web for information about Daniel Radcliffe only at some very specific times. What we need to identify is what constitutes A, T and C. The first component of Dedalo uses the Linked Data background knowledge to derive some plausible explanantia of type A to the event P. The process consists in finding facts in Linked Data that are statistically highly correlated with the explanandum. For example, both "people search for Daniel Radcliffe because a Harry Potter movie is released" and “people search for Daniel Radcliffe because the Under- 20 Football World Cup is on” are correlated with Daniel Radcliffe according to Linked Data information. If no connection between the A events and P is identified, however, there is no way to distinguish whether the A and P are happening together by coincidence (as in the second case). Also, the explanation is incomplete (and impossible) if one has no knowledge of the context C (what is the relation between Daniel Radcliffe and Harry Potter). (8)

Dedalo also finds this relationship by searching for a "Linked Data path".

The last sentence is, I think, pretty much how the ODP is supposed to help assess the quality of an explanation in a framework like Dedalo. In this case they found that Dedalo was missing a relation to a theory underlying the relationship between events (e.g. a relation to "laws"). Adding this feature is required to make the Dedalo ramework "complete" according to the explanation ODP.