Intelligent dialogue systems are an important research direction in the field of artificial intelligence， in which natural language generation technology plays a crucial role. The article comprehensively summarizes the mainstream methods of natural language generation in intelligent dialogue systems， including retrieval based and generation based methods， as well as the fusion of the two. On the basis of in depth analysis of various methods， an improved fusion based natural language generation framework is proposed， and its effectiveness is verified through experiments. Research has shown that fusion methods can integrate the advantages of retrieval and generation， improve response diversity and user interaction experience while ensuring response quality， and provide new ideas for further improvement of intelligent dialogue systems.