Research On Open Domain Text Dialogue Model For Role Consistenc

Open-domain dialogue system is an important task of natural language processing,with potential future in industries such as healthcare,finance,and the internet.The aim of open domain dialogue systems is to simulate human speaking style to communicate fluently with users.This research field has become a hot topic in recent years,remaining at the forefront of artificial intelligence research.An effective dialogue system not only requires the ability to engage in fluent and diverse conversations but also needs to exhibit similarity to human behavior in terms of personas,emotions,and styles.Maintaining stable personas is an important prerequisite for establishing long-term interactions with human,and a dialogue system needs to maintain persona stability in multi-turn and complex conversations.Current dialogue systems focus on maintaining one aspect of persona,style,or context,which leads to inconsistency in long-term conversations.The agent may respond with inconsistent replies with given personas and rigid responses that are inconsistent with the persona and style implied in the historical information and context.To address the issues in current work,such as the lack of style information in persona modeling and inconsistent personas caused by different persona distributions,this thesis focuses on the persona information and speaking style of speaker in open-domain dialogue system and studies a persona consistency model based on speaking style perception.The main contributions of this thesis are:(1)We propose a persona selection method based on persona style that provides the model with the ability to perceive speaking style.To build the persona selector,the dataset is expanded and used natural language reasoning tasks to enhance the model’s logic and generalization.A dialogue model is designed with consistent persona style using the persona selector as the main module to focus on consistency at a higher level.(2)We propose a persona-consistent model based on speaking style perception to enhance semantic and stylistic consistency in model with different persona distributions.The model is able to perceive persona speaking style through the persona selector and can adapt well to contexts with different persona distributions.During the decoding process,a generating rewrites method is used to ensure persona style consistency and persona semantic consistency.Additionally,the attention-based fusion method has been improved to adjust the influence of different information and reduce the frequency of generating persona-irrelevant responses.This thesis conducted extensive experiments on datasets with different persona distributions,verifying the effectiveness of the proposed model.The model proposed in this thesis can effectively alleviate the problems of speaking style and semantic inconsistency in corpuses with different persona distributions,thus generating fluent,diverse,and persona-consistent high-quality responses.