Learning Disentangled Semantic Spaces of Explanations via Invertible Neural Networks