Research And Application Of Open Domain Persona Dialogue Generation Method

Among the many tedious natural language processing tasks,building personalized dialogue systems is a challenging task in the field of human-machine dialogue.The open-domain dialogue systems communicate with users without any limited topic or clear goal,which can provide users with a more natural and friendly human-machine interaction experience.It has an important practical application value and has become a research topic of extensive attention by many researchers in recent years.Current open-domain dialogue generation systems are mainly based on sequence-to-sequence models,but the standard sequence-to-sequence models still suffer from generating trivial and generic responses,and the dialogue content lacking consistent personality.To address such problems,the researchers conducted in-depth studies to make the responses generated by the model more personalized.However,the existing personalized dialogue generation model still has some shortcomings,mainly in the following two points:(1)The current dialogue generation does not combine dialogue scenarios,but instead fuses persona information and dialogue context directly for response generation.(2)The limited description of personas makes it difficult for the model to understand and associate the user characteristics.The model is difficult to generate personalized response with rich information.Based on the above problems,this article takes two aspects of combining dialogue scenarios for response generation and enriching persona information descriptors as the entry point to study the personalized response generation method for open-domain dialogue systems.The main research work is divided into the following two aspects:(1)We propose an open-domain dialogue generation method based on dynamic persona fusion mechanism.Firstly,based on pre-training language model,the semantic features of dialogue text and persona information are extracted separately using the attention mechanism to learn the relationship and internal features between them.Secondly,a dynamic persona fusion mechanism is designed in the decoding stage,which can mine the relevance of dialogue context and personas according to dialogue scenarios.It can automatically adjust the weight of characteristic information in dialogue generation,so as to increase or decrease the influence of persona information on response generation.Finally,according to the extended multi-head attention mechanism,the persona information,the dialogue context and the previously decoded output word are integrated respectively,so that they jointly affect the generation of dialogue.The experimental results show that this method can flexibly fuse personas into the dialogue response according to the current dialogue scene,which verifies the effectiveness of the open-domain dialogue generation method based on the dynamic persona fusion mechanism.(2)We propose an open-domain dialogue generation method based on implicit persona information.Firstly,a personalized language model is designed to learn the user’s implicit user information from the large-scale dialogue history responses.Secondly,based on the current dialogue text,a key-value memory network is established,and a persona-aware personalized selection module is constructed to select historical responses related to the current conversation.Finally,a responses decoder is designed to fuse the learned user profiles with the selected historical responses.The personalized decoder can switch between generating a word from a generic vocabulary and copying a word from the user’s personalized vocabulary.The experimental results show that the method can reasonably utilize the dialogue history to improve the personalization degree of response generation,which verifies the effectiveness of the open-domain dialogue generation method based on the implicit persona information.(3)On the basis of the dialogue generation model proposed in this circle,an open-domain dialogue system based on persona information is designed and implemented.The whole system can realize the function of dialogue,and the personalized features exhibited by the response statements maintain semantic consistency with those in the persona information,further verifying the effectiveness of the method proposed in this circle.The development process of the system is as follows: firstly,this system pre-processes the collected data.Secondly,it loads the dialogue generation model to process the user’s input to get the dialogue response.Finally,it converts the dialogue text into speech for output.The experiment shows that the system achieve the expected effect in all functions,and the responses generated by the system improve the persona consistency.