"Mama Always Had a Way of Explaining Things So I Could Understand'': A Dialogue Corpus for Learning to Construct Explanations

Explanations aren't one-size-fits-all: The research challenges the common assumption in AI explainability that there is "one ideal explanation to be found." Instead, effective explanations are tailored through dialogue between the explainer and the specific person receiving the explanation, meaning AI systems need to adapt their explanations based on individual needs and understanding levels.

Real dialogue matters more than static explanations: The study created a corpus of 65 real video dialogues where experts explained 13 different topics to people at five different proficiency levels. This reveals that human explanation is an interactive, back-and-forth process rather than a one-directional delivery of information—a pattern that current AI explainability systems rarely capture.

Context and audience proficiency significantly shape explanations: The researchers found measurable linguistic differences in how experts explained the same topics to beginners versus advanced learners across 1,550 dialogue turns. This suggests that responsible AI systems should recognize and respond to what users already know rather than providing generic explanations.

AI can learn conversational explanation patterns: BERT-based models successfully predicted explanation topics, dialogue acts, and explanation moves using sequence information, indicating that machine learning can be trained to recognize and potentially replicate the dynamic, context-aware patterns humans use when explaining complex topics.

This work enables more human-centered AI explanations: By providing the first large-scale corpus of natural explanatory dialogues with detailed linguistic annotations, this research creates a foundation for developing AI systems that explain themselves more like humans do—responsively, adaptively, and in conversation with users.