Tackling Natural Language Inference with a logic-based method is becoming\nless and less common. While this might have been counterintuitive several\ndecades ago, nowadays it seems pretty obvious. The main reasons for such a\nconception are that (a) logic-based methods are usually brittle when it comes\nto processing wide-coverage texts, and (b) instead of automatically learning\nfrom data, they require much of manual effort for development. We make a step\ntowards to overcome such shortcomings by modeling learning from data as\nabduction: reversing a theorem-proving procedure to abduce semantic relations\nthat serve as the best explanation for the gold label of an inference problem.\nIn other words, instead of proving sentence-level inference relations with the\nhelp of lexical relations, the lexical relations are proved taking into account\nthe sentence-level inference relations. We implement the learning method in a\ntableau theorem prover for natural language and show that it improves the\nperformance of the theorem prover on the SICK dataset by 1.4% while still\nmaintaining high precision (>94%). The obtained results are competitive with\nthe state of the art among logic-based systems.\n