Research On Key Technologies Of Open-Domain Persona-Based Dialogue Systems

One of the longest-running goals in Artificial Intelligence is to make machines talk as naturally as humans.Open-domain dialogue system aims to abstract the computer into a “person”,and provide users with access to computing capabilities by the means of dialogue,so as to realize natural human-computer interaction.Nowadays the dialogue system has been widely used in real life,such as intelligent customer service systems,mobile device personal assistants,and digital humans.The dialogue systems have already been able to effectively finish the basic information-seeking tasks.However,these systems are still far from the AI vision of natural dialogue,and one of the important reasons is that these dialogue systems lack consistent personalities as real human beings.This thesis studies open-domain personalized dialogues,and its goal is to make the dialogue system present unique and consistent persona characteristics in the dialogue process.To facilitate this research,the persona information in this thesis is explicitly provided: it can be a structured key-value pair,such as “Occupation: Engineer”;It can also be a text description,such as “I am an engineer”.The key technologies needed to build an open-domain personalized dialogue system mainly focus on two aspects: dialogue generation and dialogue understanding.In the scenario of open-domain personalized dialogues,dialogue generation refers to generating a dialogue response that is compatible with the context and reasonably expresses persona information.Dialogue understanding refers to identifying the potential information in the dialogues.For example,“Welcome to Be?ing” implies the speaker is in “Be?ing”.The existing dialogue systems imitate the corpus from different sources to generate dialogue responses,which leads to inconsistency;At the same time,the lack of dialogue understanding makes it difficult to optimize or postprocess the dialogue response,which further aggravates the inconsistency in the opendomain dialogues.To address the above issues,this thesis has organized four research plans based on the generation and understanding of dialogues:Generating diversified responses with persona information: To carry out personalized dialogues,the systems need to be able to express the persona information in the dialogue responses.This thesis proposes a generation model based on the memory network and latent variable,which can model and flexibly express given persona information in conversation responses.Experiments on Persona Chat,a personalized dialogue dataset,show that this method can significantly improve the diversity of responses by expressing different persona information.Consistency understanding between persona and response: In addition to expressing persona information,it is also important to accurately understand the consistent relationship between responses and personas.In order to effectively understand the persona consistency of dialogue responses,this thesis constructs a large-scale humanannotated dataset and proposes a corresponding consistency understanding model.This thesis further verifies the effectiveness of the consistency understanding ability in downstream dialogue tasks.The experimental results show that the ability of consistent understanding can be effectively applied to the selection of dialogue responses and the evaluation of response quality.Personalized dialogue generation with consistent understanding: After being able to express persona information and understand persona consistency,this thesis further explores how to explore consistency understanding to improve persona consistency in generated responses.This thesis first introduces consistency training signals through reinforcement learning in the training stage and then leverages consistency understanding in the post-processing of responses to conduct rewriting.The experiment on the Persona Chat dataset shows that the introduction of consistency understanding can effectively improve the persona consistency of the dialogue generation model.Personalized dialogue generation under low data resource scenarios: The above studies are conducted on the premise of sufficient labeled data,but it is difficult to expand the scale of such data in practical applications.To address this issue,this thesis proposes a stacked framework of pre-trained language models that can use non-conversational data to improve the performance of the persona-based dialogue tasks,which significantly improves the capability to conduct personalized dialogues under low-resource scenarios.In general,this thesis aims to enable the open-domain dialogue system to understand and generate dialogue responses with consistent personas.For the understanding of persona consistency,this thesis explores the construction of data resources and consistency understanding method and further offers an optimization scheme for the low resource scenario.For the generation of persona-consistent dialogues,this thesis studies how to generate more diversified dialogue responses by expressing different persona during the training process and how to optimize consistency at different stages of response generation.The research in this paper can significantly improve the persona consistency of the opendomain dialogue system,resulting in more personalized dialogue responses,and further improve the user experience in various dialogue scenarios.The presented techniques also have broad application prospects in different scenarios,such as intelligent assistants,digital humans,and so on.