Research On Diverse Response Generation For Open-domain Dialog Systems

In recent years,open-domain dialog systems based on neural generation models have received increasing attention due to the availability of large-scale conversation data and the successful application of deep learning technology in dialog systems.Early open-domain dialog systems were mainly based on rules.Such rule-based dialog systems are not scalable and require a lot of manpower to write rules.While neural dialog generation models can make full use of the large-scale conversation data to learn meaningful feature representation and response generation strategies,without writing a lot of manual rulesAlthough the neural dialog generation models have achieved great success in the open-domain dialog tasks,they also face some new challenges.One of the most com-mon issues is the generic response problem.Based on the existing research,we find that there are three important factors which make it easy for the models to generate generic responses:(1)Existing neural dialog generation models are typically trained to build a one-to-one mapping from the input information to the output response.However,in practical,the relationship between the input information and the output response is one-to-many;(2)The responses in most dialog dataset are extremely unevenly dis-tributed,and the existing models are typically optimized using maximum likelihood estimation(MLE),which leads to the model's tendency to learn some high-frequency common responses;(3)The existing dialog generation models lack the ability to cap-ture the user's intention and the global structure of the generated response,which makes the models unable to understand the user's deep intention and easy to generate the generic responses that are irrelevant and monotoneThis dissertation mainly focus on open-domain dialog generation task and the main purpose of the dissertation is to solve the generic response problem in the neural dialog generation models and imporve the diversity and informativeness of the gener-ated response.Based on the above observations,the research contents of this thesis mainly include:(1)we propose a new dialog generation model based on reinforcement learning,which directly considers an given input message with multiple plausible re-sponses jointly for the one-to-many property of conversation data;(2)we introduce a discrete and interpetable latent variable in a CVAE model and solve the high frequency response problem caused by the use of the MLE method in the current neural dialog generation models by optimizing the data log-likelihood lower bound;(3)we introduce sentence functions which can reflect a user's purpose,into a CVAE-based dialog gener-ation model,so that the model can identify the user's intention and control the global structure of the generated response.