Explaining Graph Neural Networks with Large Language Models: A Counterfactual Perspective on Molecule Graphs F i n d i n g s - E M N L P