Context(s) for contextualizing explanaitons

One could argue that context is basically a setting or a situation that needs to be properly described.  Recent discussions on future XAI systems call to regard context to provide more relevant explanations. According to Sanneman and Shah, “it is not equally valuable to provide just any information to human users via XAI, but only information that is relevant to them given their respective tasks and context”.  However, the notion “context” is difficult to grasp. In fact, across disciplines, it has been recognized as a “black hole”  yielding more and more aspects as the investigation progresses. In addition to it, an interdisciplinary meaning of this notion requires two contrary positions to be integrated:  On the one hand, and considered from human sciences, context is always a product of a social construction because humans’ perception of a given situation is going hand in hand with an interpretation of it. On the other hand, the nature of artificial systems requires a context to be encoded and represented regarding users, their tasks, circumstances, and so on.  For example, a given situation – while it can be perceived individually by a human having specific experiences and associating some aspects of it with their knowledge – can be represented by a situational model. It might consist of a location at which a user will stand. This location would then set up further actions for an artificial system (ibid). However, in a case of mobile settings, in which a human is expected to move around, the situation will be more dynamic requiring a system to react to it. For the system’s application, it means that some adjustments need to be done while executing it. This simple example highlights two important aspects of a given situation: First, it is likely to change but – and second – systems need to foresee the parameters of change to adjust appropriately.

Certainly, there have been many attempts to characterize possible factors that constitute a situation.  They all require, however, that a situation can be foreseen. Because it would require the anticipation of factors affecting any concrete situation, this claim seems to be impossible to fulfill. Even more, because of widespread use of AI systems, the growing ubiquity of XAI means that users, tasks, and circumstances can change. To make an explanation relevant, thus, requires an XAI system to adapt to the changes. The capability of a system to adapt to situational changes is at the core of our systematization of “context.”

Instead of proposing lists of factors that constitute a situation, following,  we propose to specify what kind of a context the XAI system is able to observe and construct. In accordance to it, we can narrow down factors that can influence this type of context.

For our purpose, in the following, we differentiate between four types of context, as proposed in: 

Selected Context. We define this type as consisting of a selected set of material and social facts. Material facts can be a number of people, whereas their background is an example of the social facts. In fact, “context” and “situation” are often used interchangeably, but in this type, it is important to note that the material and social facts are static and preselected for an XAI system.

Situated Context. In contrast to artificial systems, humans construct a situation by interpreting it against the background of their experience. The involvement of memory is also responsible for some facts becoming constructed (e.g., persons feeling time pressure) even though they were not part of a situation in the first place.

Adjusted Context. Important to this type are possible changes to the context emerging in an unfolding interaction. In contrast to the situated context, the set of facts get adjusted for its relevance. Thus, whereas at the beginning of interaction, the set was selected and situated, the course of interaction yields some adjustments necessary to it.

Co-constructed Context. We define this type as consisting of material and social facts that emerge from the interaction. Thus, both partners contribute to the common set of facts and context is a joint achievement. In addition, the interaction itself becomes context in the sense of a background against which further knowledge is built.

The context types vary in terms of their increasing adaptability: Whereas a quite static context (e.g., for an XAI with a predefined task for a specific type of a user) consists of preselected social and material facts, these can emerge dynamically in a highly adaptive context created by the interaction between partners. The advantage of an adaptive context is to be flexible to changes of task, user, and circumstances. Whereas we divided “context” into four types, we conclude that for social XAI, all types are necessary and can be considered as building upon each other.