Eliciting Topic Hierarchies from Large Language Models

Finding topics to write about can be a mentally demanding process. However, topic hierarchies can help writers explore topics of varying levels of specificity. In this paper, we use large language models (LLMs) to help construct topic hierarchies. Although LLMs have access to such knowledge, it can be difficult to elicit due to issues of specificity, scope, and repetition. We designed and tested three different prompting techniques to find one that maximized accuracy. We found that prepending the general topic area to a prompt yielded the most accurate results with 85% accuracy. We discuss applications of this research including STEM writing, education, and content creation.