A BFO-based Ontology of Context for Social XAI

Miller, exploring explainable AI from a social science perspective, extracts four key desiderata for explanations from the literature: explanations are contrastive, selective, social, and causal rather than statistical.  He argues that these factors are essential to consider for developing truly explainable AI systems that offer explanations. He further argues that all these desiderata converge around a single point: explanations are contextual. Yet, despite the contextual nature of explanations, the XAI community tends “to abstract from the social and political embeddedness of AI systems”.  Thommes also discusses that focusing on one-fits-all solutions and disregarding the contextual factors of explanations (e.g., the explanatory needs of explainees) limits the usability of XAI systems.  In line with these concerns, Rohlfing et al. have proposed a social design of XAI, in which explainers and explainees co-construct an explanation.  They have further introduced the fundamentals of the new paradigm, Social XAI, grounded in social interaction research.  

In this work, we aim to propose formalizations of different types of context to make this concept more accessible for future XAI development. To this aim, we use Basic Formal Ontology (BFO), an ISO-standard top-level ontology.  BFO is domain-neutral, provides semantic interoperability among BFO-based domain ontologies, and is state-of-the-art, widely used, and well-documented. Following the basic types of entities in BFO,  we divide the entities relevant to context into continuants (entities that continue or persist through time, such as independent objects, qualities and functions) and occurrents (entities that occur or happen, such as processes or events). 

This work continues our previous efforts to formalize Social XAI concepts using BFO.  Booshehri et al  discuss the multifaceted nature of context in Social XAI, noting that definitions of conversation context in the literature highlight various modes of being – that is, they refer to entities that are either continuants or occurrents. However, for the sake of conceptual clarity and practical utility, it is advisable to focus on a single mode of being at a time. 

In this vein, Booshehri et al.  have discussed classifications for different types of context as follows. One such type is spatial context, which is “the site of the conversation, formed by the physical place [. . . ] in which the conversation takes place.”  Spatial context can be classified as a BFO: site, which is, “intuitively, an immaterial entity in which objects [. . . ] are or can be contained”.  Another important type is co-constructed context, introduced by Rohlfing et al.,  and defined as being “a dynamic set of facts that is not set up beforehand the interaction but emerges from it incrementally and locally; the emerging joint knowledge provides a context for further actions.” In this sense, a co-constructed context can be modeled as an aggregate of generically dependent continuants (GDCs) with each fact being considered as a GDC (copyable patterns that exist independently of a specific bearer as long as some bearer exists; facts are added to or removed from GDC aggregates as they become relevant or irrelevant). A third type refers to the processual aspect of context following Landgrebe and Smith  who state that “[t]he conversation is its own context at all levels of language production and interpretation.” In this sense, one could consider context as being a BFO: process. Recognizing this processual nature enables modeling the context of an explanation across different modalities and to cognitively profile different measurable dimensions of context using the notion of process profile in BFO (not yet part of the ISO standard, but see here ). An example of this would be verbal context, where one abstracts away from other (non-verbal) modalities – such as facial expression or eye gaze – to isolate and examine one dimension of interaction, such as speech acts. 

We will present ongoing work towards classifying additional types of context that are relevant to Social XAI (introduced in Rohlfing et al. ) within the standardized framework of Basic Formal Ontology. Our goal is to develop a BFO-based ontology of context for Social XAI that contributes to a broader ontology development effort focused on an ontology of explanation with applications to Social XAI.