Make a hierarchical pipeline for currating SVGs to my data needs

I need a pipeline that does what ReGroup: Recursive Neural Networks for Hierarchical Grouping of Vector Graphic Primitives does semantically. I want a pipeline that given a purely human made svg where semantic categories are obvious, no gradient, no opacity, no 3d effects, etc... we can at an accuracy that fools me make a dataset like the one provided. I am happy to answer further questions too. I absolutely do not want to resort to man made curation. I have an enormous dataset that I can test whether your pipeline will work or not. You are allowed to use a closed source AI as long as it is reachable through an API and will be.