Towards Linguistically Informed Multi-Objective Pre-Training for Natural Language Inference