Beyond One-Size-Fits-All: Inversion Learning for Highly Effective NLG Evaluation Prompts