Research On Personalized Conversation Generation Based On User Portrait

Nowadays,human-machine dialogue systems have received increasing attention.The task of dialogue generation has always been one of the most popular research directions in natural language processing.However,mainstream human-machine dialogue systems rarely consider the personalized features of the speaker.An important and unexplored aspect of dialogue systems is to enhance the quality of personalized dialogue responses based on the user profile of the interacting parties.Personalization is the key to creating intelligent dialogue systems that can be most effectively adapted to human life.However,it is difficult to reflect personality in natural language processing.How to effectively use user profiles to improve the quality of personalized dialogue has become a hot research direction and a key challenge in dialogue generation tasks.Traditional dialogue generation methods ofien only extract text information features to generate dialogue,while ignoring user profile information such as personality,gender,and occupation.This article mainly proposes three effective solutions for personalized dialogue generation based on the above shortcomings:personalized dialogue generation combined with user profile emotion information,personalized dialogue generation based on user profile and pre-trained models,and personalized dialogue generation based on user profile and template construction.Experimental analysis shows that the proposed solutions in this article have significant improvements over traditional personalized dialogue generation methods.Specifically,the work in this article mainly includes the following three points:Firstly,to address the issue that traditional personalized dialogue generation methods have not utilized emotional information from user profiles,a personalized dialogue generation method that incorporates emotional information from user profiles is proposed.This method uses a BERT-MRC model to extract emotional and attribute information from character profiles and historical conversations.An improved UNILM neural network model is used to encode the personality of characters and historical conversations,combining emotional and attribute information during the encoding process to generate dialogues that are consistent with the character’s personality.Experimental results show that incorporating emotional information improves the quality of personalized dialogue generation and increases the diversity of generated responses.Secondly,to address the issue that traditional personalized dialogue generation methods cannot express the relationship between characters,a personalized dialogue generation method based on user profiles and pre-trained models is proposed.Specifically,this article proposes a personalized relationship tree structure to represent the personality and relationships between characters in a tree structure.During the pre-training phase,the BART model predicts the personalized relationship tree to learn the speaker’s personality and the relationships between characters,and transfers the learned knowledge to the dialogue generation phase.Experimental results show that the proposed model significantly improves evaluation metrics such as BLEU and PPL,and can effectively improve the personalized performance of dialogues.The research also shows the importance of the personalized relationship tree structure to personalized dialogue models.Finally,as the traditional personalized dialogue generation method can not naturally integrate the Semantic information and dialogue content of personalized structure with the rich knowledge of the pre training model in the form of natural sentences,a personalized dialogue generation method based on user portraits and prompt learning is proposed.Specifically,this article first constructs a dialogue template that combines user profiles,giving speakers with personal attributes and relationships,as well as historical dialogues corresponding to the speakers.Traditional dialogues and character information are concatenated to form a smooth natural language expression.Then,in the micro tuning stage,prompt learning is used to output user profile information and personalized dialogues,Enable the model to naturally learn the relationship between user profiles and conversations.The experimental results show that the model proposed in this article has significant improvements in indicators such as BLEU,which can greatly enhance the personalized performance of conversations.