An Open Domain Question Answering System Trained by Reinforcement Learning