Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks’ Internal Representations