Adaptive Explanations as Co‐Constructed Processes. Modeling a Rational Explainer Through the Interaction of Dynamic Partner Models and Non‐Stationary Decision Making

In explanations, humans consistently adapt to their conversation partners. Such an adaptive explanation is a complex and highly dynamic process. This raises importantquestions: How do humans adapt their explanation in such a short time despite having limited and incomplete knowledge about the interlocutor? How do changes in the mental representation of the interlocutor manifest in verbal behavior patterns? And considering the increasing presence of artificial agents in our daily lives, (how) can those patterns be transferred to human-agent explanations? This thesis presents an abstract framework for modeling a human-inspired adaptiveexplainer along with a publicly available implementation of it. This sequential non-stationary decision process model for adaptive explanation generation using partnermodels (SNAPE-PM) collaborates with the user, making the explanation a coconstructed product. The development of such a rational agent needs to fulfill three key requirements: (1) it must facilitate **adaptive** interactions, (2) the decision-making process should be **transparent** to understand the underlying cognitive or linguistic processes and their effects, and (3) the evaluation should be **comparable** to other approaches, enhancing the generalizability of results. The agent’s modular Architecture utilizes Bayesian and Markovian approaches for transparent partner-modeling and decision-making, coupled with generative models for both language perception and production. By capturing the user’s knowledge along with other adaptation parameters like attentiveness and expertise, SNAPE-PM can transfer adaptation strategies from human-human explanations. Acknowledging that a partner model evolves not just with the user’s knowledge but also with other pertinent factors is a unique feature of this approach. Finally, the thesis evaluates the performance of the agent in a series of online user studies. Results indicate that an interaction with SNAPE-PM results in a significantly higher understanding in comparison to a non-adaptive explanation. Notably, aligning with the user’s knowledge only did not yield meaningful differences. By highlighting the methodological questions and challenges that arise when using established methods to assess the performance of adaptive interactions, comparability issues of adaptive approaches are also addressed. This doctoral thesis examines adaptive explanatory processes from an interdisciplinary perspective acknowledging the diversity and flexibility of a human Partner model. The focus is not on explainable, but on explanatory artificial intelligence, by which findings and methods from linguistics, psychology, cognitive science, and Computer science are brought together in a rational agent.