Skip to main content Skip to navigation

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

Project Overview

The document explores the role of generative AI in education, particularly through the development of a corpus designed to enhance dialogical explanations in explainable AI (XAI). It highlights the necessity for AI systems to communicate concepts effectively in a conversational format, tailored to the diverse proficiency levels of learners, ranging from children to professionals. This corpus includes 65 transcribed dialogues from the '5 Levels' video series, meticulously annotated for dialogue acts, topics, and explanation moves. Key findings reveal that human explanations are collaborative, evolving through interaction, and that their effectiveness is significantly influenced by the explainee's prior knowledge. The insights gleaned from these dialogues underscore the potential for generative AI to facilitate personalized learning experiences, improving comprehension and engagement across various educational contexts. By focusing on the nuances of dialogical interaction, the research contributes to the development of AI systems that can better support learners by providing contextually relevant and comprehensible explanations.

Key Applications

Dialogue corpus for learning to construct explanations

Context: Educational context involving varying proficiency levels (children to colleagues) in understanding complex topics like blockchain and machine learning.

Implementation: Annotated corpus created from dialogues in the '5 Levels' video series, focusing on how explanations are constructed in dialogues based on the explainee's proficiency.

Outcomes: Insights into how experts tailor explanations to different audiences and baseline results indicate that modeling sequential dialogue interaction helps predict topics and dialogue acts effectively.

Challenges: Limited corpus size restricts deeper statistical analyses; data sparsity affects the predictability of certain labels.

Implementation Barriers

Data-related barrier

The limited size of the corpus prevents comprehensive statistical analysis and effective training of dialogue systems.

Proposed Solutions: Future work should target increasing the scale and heterogeneity of explaining data to enhance the effectiveness of AI systems.

Project Team

Henning Wachsmuth

Researcher

Milad Alshomary

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Henning Wachsmuth, Milad Alshomary

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies