Modeling the Quality of Dialogical Explanations

Expert explainers usually plan an explanation strategy by choosing appropriate explanation moves, dialogue acts, and topics to ensure optimal comprehension on the explainee side (Wachsmuth and Alshomary, 2022). Additionally, explainees may actively engage in dialogues by asking clarification questions and providing feedback to ensure they understand the information correctly (Madumal et al., 2019).

In this work, we take a first substantial step towards studying the quality of daily-life explanation dialogues concerning the explainee’s understanding. We hypothesize that the interactions in explanation dialogues in terms of explanation moves, dialogue acts, and topics correlate with explanation success. To study this hypothesis, we construct the first corpus of daily-life explanation dialogues. We then compare this corpus to existing expert explanation dialogues (Wachsmuth and Alshomary, 2022), and we evaluate the effectiveness of pretrained language models in predicting the quality of explanation dialogues.

created corpus consists of 399 daily-life explanation dialogues from the Reddit forum “Explain Like I am Five (ELI5)”. One example dialogue is shown in Figure 1. We annotate the corpus for the explanation quality of each dialogue as well as the interaction concepts of Wachsmuth and Alshomary (2022) (Section 3)

Rohlfing et al. (2021) build on this view, clarifying that explaining in an intrinsically dialogical process in which the participants co-construct an explanation. They highlight the importance of successful communication between the explainer and explainee, which is a challenge that research needs to address adequately.