Cognitive and Interactive Adaptivity to the Explainee in an Explanatory Dialogue: An Experimental Study

Depending on the partners, a different dialogue evolves. When focusing on the dialogue as a product, partners can be considered as a context factor impacting this product. In the literature, this effect is explained by adaptation processes taking place between the partners.  Even more pronounced effects can be observed in asymmetric interactions. Explanatory dialogues are asymmetric because the explainer (EX) knows more than the explainee (EE).  Even though we have evidence that partners adapt to each other, we know little about how the adaptation takes place. In this paper, we consider two mechanisms that are intertwined with each other and drive the adaptation of the partners. 

The cognitive adaptivity refers to the adaptation of the partner model (PM), i.e., the mental representation that the EX has of the EE. PM-dimensions can be differentiated in dispositional dimensions beyond the specific situation, and situational dimensions [3, p. 230], that are valid in the specific situation. Before a dialogue, the EX has a global PM formed by prior experiences with similar EEs, stereotypes etc..  Assuming that a dialogue is a process that unfolds iteratively,  we propose that the PMs evolve and change as the dialogue progresses.  The partners may vary in their knowledge or interest in the explanandum. 

The interactive adaptivity, in turn, refers to the adaptation of verbal behaviors. In this vein, the literature suggests mechanisms such as alignment or interpersonal synergy, according to which partners align on various linguistic levels.  We consider the level of speaker moves by the EX. Chi et al.  refer to speaking strategies as speaker moves. The move diversity describes the number of different moves which are used by a speaker. Thereby, it shows their repertoire of different explanation moves. For the overall adaptation of the partners, we further assume that the cognitive adaptivity influences the EX’s verbal behavior. A less interested EE and the resulting mental representation of this partner may lead to a shorter explanation. Based on Buhl et al.,  experimental settings are needed to control one partner's behavior and examine its influence on the other partner to further understand the intertwinement of cognitive and interactive adaptivity as a driving force. Thus, we raise the following research questions: How do dimensions of the EXs’ PM change during the explanation, depending on EEs’ verbal behaviors? How does the move diversity of the EX depend on EEs’ verbal behaviors? 

To answer these questions, a role-playing experiment was conducted. A confederate took the role of the EE, to whom a participant explained a board game. Our experiment is based on two conditions: In the Passive Condition, the confederate (EE) was instructed to only use neutral backchannels (“mmh”), and no other signals of mood orunderstanding. In the second Active Condition, three EEs were trained to be co-constructive. Their behaviors consisted of backchannels as well as speaker moves that were found to relate to changes in the PM,  such as factual question, paraphrasing partner, summarizing info and example (see OSF for speaker moves). The explanations in the passive condition lasted on average 2.5 minutes (SD = 0.91 min.). In the active condition they were longer with an average of 6.52 minutes (SD = 2.27). A total number of 84 participants took part in the experiment. They were predominantly university students (age M = 23.42 years, SD = 3.51, 51 women, 32 men, 1 diverse).The experiment consisted of three phases: Before the explanation, the EXs were asked about their PM, then they explained, and afterwards they were asked again about their PM. 

The PM was assessed with a standardized questionnaire regarding twelve dimensions. For this analysis, we focused on three dispositional dimensions (dispositional knowledge, interest in explanandum, and communion), and three situational dimensions (situational knowledge, interest in the explanation and co-construction). The move diversity describes the number of different moves which are used by a speaker. Thereby, it showcases their usage of different explanation moves. 

To answer the research question about changes of PM-dimensions during the explanation, depending on EEs’ verbal behavior, 2 x 2 ANOVAs were conducted with two measurement points (before vs. after the explanation) and two conditions (passive vs. active EE). The interaction between the measurement point and the condition is informative. All situational dimensions exhibited significant interaction effects (p < .05), whereas none of the dispositional dimensions showed significant interactions. 

The most frequent EX moves were providing info, additional info, paraphrasing both, and paraphrasing self, all of which occurred significantly more often in the active than passive condition. In the passive condition, the EXs disposed of a repertoire of on average 5.2 different moves (SD = 1.4) and in the active condition, they had an average of 9.35 (SD = 2). Comparing both conditions, the average move diversity is significantly less (p < .05) in the passive condition. 

An experiment with extreme condition variation investigated how the EE's behavior influences the EX’s PM and explanations. The changes in the PM confirm the assumption that it is influenced by the behaviors of the EE. In this case, changes were mainly found in situational dimensions, not in dispositional dimensions. However, this does not rule out the possibility that these or other dispositional PM-dimensions can change during a dialogue. An example would be the EEs stating that they have no prior knowledge of games or that they dislike them. The EX moves were analyzed regarding the move diversity, which depended on the EE's verbal behavior. Limiting this behavior reduced move diversity and thereby one aspect of adaptivity. 

For explainable AI, the question of relevant dynamic dimensions of the PM and their influence on the explanation arises to personalize explanations.  From these results, indications of the interplay of cognitive adaptivity of the PM and interactive adaptivity in the explanations can be derived. 

On the basis of this investigation, in next steps, the effect of the explanation on the EE should be investigated. Additionally, adaptivity should be analyzed in the situational interplay of moves and PM-dimensions in sequences of the explanation, also in AI—human interactions.