Improving Information Extraction by Acquiring External Evidence with Reinforcement Learning