Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset