| The internet contains an ever-increasing body of human communication. The emerging class of pre-trained generative language models represents a new opportunity for ingesting, understanding, and summarizing these communications. This work outlines a method of generating Prototype Relation networks using a pipeline of large language model prompt-completions. These networks are constructed with nodes of prototypical authors that have their views represented within the corpus, and edges containing possible argumentative relations between them. Methods of evaluating these networks are described, and are used to show that the content within the generated prototype descriptions and relation descriptions are in line with observations of hold-out sets and synthetically constructed comparisons. The pipeline and evaluation metrics make use of a prompt-based approach to argument role labeling, which is also tested, showing current generations of models can reach argument labeling accuracies on par with baselines. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.