A dialog-based approach to explaining machine learning models (B01)

与目前的算法透明模型假设一个纯粹的算法解释的建设，项目B01采用了一个共同建设的方法，解释者和被解释者之间的互动和特定领域的过程。我们的目标是引出和共同构建一个通用的和可实例化的对话协议，在特定的社会相关领域（健康，保险或警务）内实践。交互式问答协议将侧重于反事实问题，以增强对输入与分类器输出之间关系的理解。

In contrast to current models of algorithmic transparency assuming a purely algorithmic construction of the explanation, Project B01 adopts a co-construction approach to the interactive and domain-specific process between explainer and explainee. The goal is to elicit and co-construct a generic and instantiable dialog protocol that is practiced within a specific domain of social relevance (health, insurance, or policing). The interactive question–answering protocol will focus on counterfactual questions that enhance the understanding of how inputs relate to outputs of a classifier.