Research On Persona-Grounded Open-Domain Dialogue Generation

In open domain dialogue systems,chatbots engage in free conversations with users without specific purposes,aiming to provide users with a high-quality conversational ex-perience.Endowing chatbots with persona information(interests,professions,etc.)can help generate more human-like and informative content.However,due to the brevity of the persona and the uncertainty of the conversation flow,the model encounters certain difficulties in accurately selecting the persona,resulting in inconsistent content with the reference.Additionally,the failure to consider the impact of context on persona expression leads to insufficient coherence,and incomplete understanding of the user’s state makes it easy for the model to generate responses lacking empathy.This paper focuses on the re-search hotspot of personalized dialogue systems,namely,how to combine the persona that is most reasonable with the dialogue context,and generate higher quality responses based on the persona and dialogue context.First,for the unstructured persona descriptions in open domain dialogue systems,this paper proposes a dialogue generation model based on reasonable persona utilization considering conversation flows.The model selects the most appropriate persona by keep-ing track of the historical persona transitions as well as leveraging the reference response information as guiding signals during training,to generate target responses.Experiments demonstrate that this method can generate more consistent responses.In addition,for the appropriate persona and dialogue context,this paper proposes a disentangled-attention based framework for personalized dialogue generation.The model considers the influence of specific dialogue content on the representation of persona dur-ing response generation.It uses Transformer as the backbone to construct three kinds of decoding adapters,i.e.,persona,context,and persona-context,aiming to dynamically decode each token by extracting different information sources during the generation pro-cess,and the probability distribution of word prediction is further refined using the A*-like heuristic strategy.Experiments demonstrate that the model can generate responses with higher coherence.At last,to enhance the personalized dialogue system’s understanding of user states,this paper proposes a retrieval-augmented prompting learning framework for personalized empathetic dialogue generation.The method pre-trains a dense passage retrieval(DPR)model through contrast learning,which can fetch the exemplary responses in a real hu-man conversation corpus that best match the given persona and dialogue context,and these exemplars are then adopted as auxiliary information for target response generation in combination with heterogeneous prompt learning.Experiments demonstrate that the model can generate responses with higher empathy.Overall,the research work in this paper addresses some of the existing problems in the field of persona-grounded open domain dialogue generation in three aspects,and improves the quality of personalized conversation generation.