Neural Conversation with Memory Mechanism

We propose a neural conversation system with memory mechanism to realize natural conversation exchanges considering the previous utterances. The neural conversation system consists of a Sequence-to-Sequence model and a memory mechanism. The Sequence-to-Sequence model can generate gramatically correct replies and the memory network can consider the previous utterances. The proposed method is trained with Cornell Movie-Dialog corpus and realize conversations between human and a; computer. We confirm that the proposed method can generate replies depending on the previous utterances but it is difficult to generate semantically correct utterances.