Neural Feature-Adaptation for Symbolic Predictions Using Pre-Training and Semantic Loss