Probing Linguistic Information For Logical Inference In Pre-trained Language Models