Aligning Characteristic Descriptors with Images for Human-Expert-like Explainability