Research And Implementation Of Multi-Turn Dialogue Systems For Complex Tasks Based On Deep Learning

With the development of artificial intelligence,man-machine dialogue system has been widely used in virtual assistant,intelligent customer service and other fields.Among them,the task-oriented multi-turn dialogue system has attracted key attention in academics and industry,which solve domain-oriented specific tasks,such as weather search,attraction recommendation,hotel booking,etc.The system guide needs to help a user complete task or achieve goals.Traditionally,the interaction of taskoriented dialogue systems is often limited to a single domain,difficult to deal with the dialogue scenes involving complex tasks or spanning multiple domains,and lack of scalability.In industry,users expect that the system can not only reflect diversified personality preferences,but also be related to specific transactions and goals,that is,explore a unified solution in design and implementation for chatbots and task-based systems.To tackle the aforementioned challenges,this paper proposes a multiturn dialogue systems for complex tasks based on deep learning,oriented to the complex-task scenes,solving the problem of the lack of universality,generalization and controllability of the traditional models in multi-domain multi-turn situations.Firstly,in view of current demands for dialogue systems in the industry,this paper starts with the research in dialogue models to analyze and optimize them in terms of existing problems in multi-turn intent tracking,cross-domain switching and transition,diversity and controllability in response generation in depth.Specifically,we propose a hierarchical gate enhanced dialogue state tracking framework,which replaces the traditional multi-label classification approach with the open-vocabulary based approach,which solves the challenges of dependence on static ontology as well as redundant tracking,and improves the robustness and scalability of dialogue models.Then,this paper further proposes an end-to-end construction and management framework for general dialogue models.This method achieves the unified framework of two kinds heterogeneous tasks,chitchat and task-oriented dialogue based on a single general language model,and solves conversation interaction in practical transaction by breaking the barriers between different vertical domains with lightweight parameters and low inference time,based on pretrain-finetune,data augment,machine teaching and other skills.Apart from above dialogue research,this paper further studies the controllability of the two tasks of dialogue strategy management and system response generation in task-oriented dialogue,using a plug-and-play controller to produce human-like interactive utterances with different styles according to the style of writing,so as to improve the interactivity of dialogue and consumer experience.In the process of in-depth study of multi-turn intelligent Q&A algorithm,this paper implements a series of Chinese multi-turn dialogue systems for composite tasks with enterprise application value,which are presented in the form of web application.Through diversified data resources,the system can flexibly deal with chitchat and task-oriented dialogue,and freely control the generation of system strategy and response to improve the user experience.Some key technologies are optimized according to the characteristics and application scenarios of the projects,which verifies the industrial value of the man-machine dialogue technology in the field of natural language processing.