![](https://csdnimg.cn/release/download_crawler_static/12256631/bg4.jpg)
S. Chari, D. Gruen, O. Seneviratne, D. McGuinness /
standing of the different explanation types, can help us generate explanations more suited
to the users’ needs.
In the rest of this section, we will present a review of the survey and position papers
that we used as guides to present the explanation types.
2.1. Findings from Review and Position papers
A recent position paper [5] presents links to explanations from the social sciences domain
to the explanation needs in the explainable AI community. While this paper posits that
their ideas are focused on how to better communicate the explanations of interpretabil-
ity of black-box, deep learning models, the principles discussed apply to a broader class
of AI models. The authors suggest that everyday and scientific explanations used in do-
mains such as psychology and social sciences, are also applicable to explaining AI mod-
els because they are able to present abstract information at different granularities. Fur-
ther, they draw a parallel between scientific explanations and trace-based explanations,
in that both of these might not be understandable to all users. Hence, they present a case
that, when explanations are delivered to humans who are selective and social in their
processes, the explanations need to be contrastive and communicative. They conclude
that the explainable AI community needs to provide explanations that are directly tar-
geted and tailored to the needs of the users. We note that, in our taxonomy of explanation
types (Table 1), we don’t account for implementation challenges, and we include certain
explanations, such as everyday explanations, that require common-sense knowledge that
may be difficult to gather and operationalize.
Furthermore, another survey paper [2] motivates the need to leverage explanation
science literature from related fields, such as constraint programming, forensic sciences,
context-aware systems, case-based reasoning, causal discovery, etc. The authors present
a brief description of the usage of explanations in these domains, and use these related
fields to motivate their plausible adoption in the explanation of the interpretability of ML
models. In addition, the DARPA XAI report [24] also lists desirable explanation types,
focusing on their form of delivery (explanation modalities), including visualizations, an-
alytical statements, alternate choices, and case-based presentations.
2.2. Summary
Research in different explanation types is influenced by research from interdisciplinary
fields, spanning social sciences and philosophy. While generating our catalog of defi-
nitions for different explanation types (Table 1), we identified that certain explanation
types, such as contrastive, counterfactual, case-based, and trace-based explanations, are
well-documented in the computer science literature. However, for some other explana-
tion types, such as statistical, everyday, and scientific explanations, we had to refer to lit-
erature from other domains, such as philosophy. We believe that explanations presented
to users include components from different explanation types. For example, users require
causal justifications to trust AI systems [25], statistical evidence for future exploration,
scientific summaries to comprehend the system [26], and everyday explanations con-
ducive to their understanding [5]. Hence, we believe that there is a need to expand the
explanation types our AI systems can currently support and build explanation facilities to
generate hybrid and user-centric explanations. In Section 3, we present technologies that,