Abstract: One of the key tasks of open-domain dialog system is to generate diverse and coherent dialog responses.However,one-way inference from above information alone cannot achieve this goal.To solve this problem,this paper proposes a bidirectional inference model MLVBI(multiple latent variables bidirectional inference) based on multiple latent variables.First,variational auto-encoder is incorporated into the language model and one-way inference is extended to two-way inference.That is,after the corpus is divided into context,query and response,forward inference is used to infer the response from the query to learn the word order information,and reverse inference is used to infer the query from the response to learn additional topic information at the same time.Finally,the model is integrated into bidirectional inference to generate more coherent responses.Then,in order to solve the problem of insufficient explanation ability of a single latent variable in the two-way inference process,this paper introduces multiple latent variables in the inference process to further improve the diversity of generated conversations.Experimental results show that MLVBI obtains the best accuracy and diversity on two open-domain datasets,DailyDialog and PersonalChat,and ablation experiments also show the effectiveness of two-way inference and multiple latent variables.

Key words: Dialogue generation, Variational auto-encoder, Latent variable, Bidirectional inference, Long short-term memory