| Understanding the ebb and flow of online conversation has become a core task in a variety of domains. Public policy, public relations, marketing, and a host of other fields concern themselves with extracting, predicting, and reacting to, changes in the topics being discussed by online users, and the disposition these users have with respect to topics of interest. Creating systems that can automate or simplify this process would have an immediate effect on these endeavors. To that end, this contribution proposes a method of leveraging large language models to process corpus' of online content and comments to generate a set of descriptive profiles describing the hypothetical positions of a commentator engaged with the content. We propose a method of crafting prompts for language models that tasks them with generating these 'Ideal Profiles'. This method is used to generate profiles based on a corpus of news articles and their associated comments. To evaluate their utility, learned topic models are fit to the article and comment data, as well as manually constructed sets of comparison profiles. The learned topic models are used to evaluate perplexity and coherence metrics between the generated profiles and evaluation corpus’. This paper highlights results that suggest that the profiles generated contain meaningful topics, and that they have coherence with manually constructed profiles. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.