Towards Controllable Natural Language Inference through Lexical Inference Types