Probing Natural Language Inference Models through Semantic Fragments