Learning from Ambiguous Demonstrations with Self-Explanation Guided Reinforcement Learning