| Social media platforms have become the dominant location of discourse. This new town square is comprised of numerous collections of comments and posts. Understanding the discourse that happens within these comment-based discussions is of huge importance to anyone seeking to profit from or extract information from the internet. Large Language Models offer a new opportunity to extract this type of information. This work describes and evaluates a pipeline for generating network representations of comments, as well as an approach to using these representations to improve performance on social media popularity prediction tasks. We describe how to generate a Prototype Relation Network (PRN), how it can be used as a summary of an online conversation, and how it can be converted into useful features. This work shows that a graph-convolution-based predictive model that uses these features as input improves performance across a set of social media popularity prediction tasks. Comparison with a baseline lacking these network-derived features shows improved performance across regression, binary classification, and multi-label classification. This pattern is shown to hold across four different major language models, and three major embedding models. |
*** 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.