Natural language understanding (NLU) is an important module of spoken dialogue systems. One of the difficulties when it comes to adapting NLU to new domains is the high cost of constructing new training data for each domain. To reduce this cost, we propose a zero-shot learning of NLU that takes into account the sequential structures of sentences together with general question types across different domains. Experimental results show that our methods achieve higher accuracy than baseline methods in two completely different domains (insurance and sightseeing).