Research On Neural Network Conversation Model Based On Conversation Style Transfer

In recent years,with the increasing usage scenario,and the continuous progress of supporting technology,the number of intelligent dialogue system research is also increasing.Many scientific research institutions and enterprise organizations are trying to build a dialogue system to solve the actual problems,in which,the chatbot dialogue system is attracting people's attention because of its extensive use scenarios.The chatbot dialogue system is dedicated to the natural and smooth conversation with human beings in the unlimited domain.This paper focuses on the study of the chatbot dialogue system,and in addition,the diversity of generated responses and the quality of generated responses are improved through two ways: form style transfer and text style transfer.First of all,in the dialog generation task,a dialog generative model based on retrieved responses fusion mechanism is proposed to solve the problem of safe response in the traditional generative model.The model first uses the retrieval model to obtain the candidate Q & A pairs and uses the Long Short-Term Memory neural network structure with the fusion mechanism to fuse the retrieved responses.This structure could put the retrieved responses into every step of the generative model and improve the utilization ability of the model.Experiments are conducted in the Chinese and English data sets.The experimental results show that this method has a significant improvement over the baseline model,and can better improve the diversity and accuracy of the generated results.After that,in the task of dialog form style transfer,aiming at the problem that most of the existing conversational models can only reply through text and mainly declarative sentences.this paper proposes a dialogue generative model based on the question timing catching and a dialogue generative model based on image meme.In the former,the discriminator is constructed by using the natural tags of the multi-turn dialog data set to train the model and find the right time to ask questions.The latter uses emotion as a link to use the multimodal information in the image meme,and obtains the text embedding information and emotion embedding information of the input query and image meme respectively through different pre-trained models,and introduces the method of contrastive learning to train the model to solve the problem of large amount and scattered distribution of image meme data,and uses the image meme as the reply form of the dialogue model through the above methods.The experimental results show that the first model can effectively help multi-turn dialogue and greatly improve the number of rounds of dialogue.The second model can better select and recommend image meme,and can better complete the task of image meme reply.Finally,in the task of conversational text style transfer,aiming at the problem that the existing conversational generation model cannot be transferred to any specific style text,this paper proposes a dialog generative model based on text style transfer.The model uses a twostage generation method.First,the dialog generative model based on the retrieved responses fusion mechanism is used to generate the initial response,and then the final response of a specific style is generated through the style transfer model.In order to solve the problem that there is no specific style in initial response in the dialogue generation task,this paper introduces a reinforcement learning method and a cyclic encoding method for model training.The experimental results show that the proposed method can greatly enhance the intensity of style transfer and retain the original text information.