Multi-turn dialogue generation considering speech information of each speaker by adding recurrent neural networks

A dialogue generation method using neural networks (NNs) has been proposed. The HRED model is a model for multi-turn dialogues by creating a hierarchical structure by layering several encoder-decoder models. Furthermore, the VHRED model generates a variety of responses by adding latent variables. However, since these models do not consider the user who has spoken, they generate inconsistent responses in the same dialogue, which is a problem. In this study, instead of using the user's embedding vector, we add a user recurrent NN (User-RNN) to retain the speech information of each speaker and generate consistent responses.