Research On Ubuntu Dialogue Generation Model With Mixed Distribution Hierarchical Latent Variables Combining Topic And Emotion Information

Dialogue model has always been the important research direction in the field of artificial intelligence.Dialogue models are divided into dialogue retrieval models and dialogue generation models.The dialogue retrieval model retrieves the most appropriate reply by calculating the sentence similarity and returns it to the user,which doesn’t reflect the idea of "human-like".The dialogue generation model is completely generated by the computer according to the content of the dialogue,so the dialogue generation model is more in line with the goal of artificial intelligence.Researchers often conduct related experiments on the Ubuntu dialogue dataset.The contents of the dataset are the large number of users discussing how to use the Ubuntu system.It also contains noise information,but the previous research work didn’t start from the dataset itself to improve dialogue.This thesis uses the Ubuntu dialogue dataset as the research object,and does the research work on the generative dialogue system,which contributions are as follows:(1)Aiming at the problem that the dialogue generated by the dialogue generation model trained on the dataset doesn’t match the topic and emotion due to the general large scale and complex structure of the dialogue dataset,this thesis takes the Ubuntu dialogue dataset as the object,and combines the topic,emotion model and dialogue generation model to improve the quality of the dialogue generation model.This thesis firstly analyzes the Ubuntu dialogue dataset in detail,and then extracts and adds topic and emotion information.After screening reasonable and valuable dialogues through statistics and analysis,this thesis uses three topic models(LDA,K-means and CTM)and two emotion models(Vader,Text2Emotion)to extract the topic and emotional information of the dataset,and then the two kinds of information are spliced into five baseline models and compared with the original dataset.The results show that,compared with the results of the dataset without topic sentiment information,the five models improve all evaluation indicators by nearly 2%.(2)Aiming at the problem that the current hierarchical latent variable dialogue generation model doesn’t match the context and the posterior collapse,this thesis proposes a hierarchical latent variable dialogue generation model based on mixture distribution.First,the model uses three encoders to encode the overall dialogues,each dialogue and each utterance of dialogue containing topic and emotional information.Then we calculate the global semantic latent variables sampled by Gaussian distribution and Dirichlet distribution.The local semantic latent variables of each sentence are obtained by sampling inference based on the global semantic latent variables.Finally,the corresponding dialogues are generated according to the local semantic latent variables.In addition,this model also sets a cyclically varying KL divergence coefficient,which effectively alleviates the posterior collapse problem of the model.We did experiments on the processed Ubuntu dialogue dataset and Cornell movie dialogue dataset.The experimental results show that,compared with the existing optimal hierarchical dialogue models,the proposed model combines topic and sentiment information,is effective in all evaluations.The indicator has increased by nearly 2%.(3)This thesis designs the generative dialogue prototyping system for Ubuntu users.The system can realize the functions of dialogue generation answer,similar question recommendation,hottest question recommendation,latest question recommendation and so on.The responses generated by this system are specifically for each user who asks a question,and the responses are generated through multiple rounds of dialogues.The system can also recommend similar questions to users.There is also a list of the latest and hottest issues on the system home page,and users can view the latest and hottest issues.