Design And Implementation Of Dialogue System Based On Few-Shot Intent Detection

With the development of artificial intelligence,dialogue systems have been applied in many fields and products.Among them,task-based dialogue systems can help enterprises reduce labor costs,improve communication efficiency,and improve service quality.There are many challenges in building this system:(1)is highly dependent on massive training data;(2)business-oriented dialogue management is difficult to build based on end-to-end models;(3)it cannot effectively utilize user dialogue history data.The goal of this thesis is to design and implement a few-shot intent recognition model and a few-shot task-based dialogue system.Around this goal,this thesis starts from two aspects:the research of the few-shot intent recognition model and the system implementation and proposes a solution for the shortcomings of the existing technology.To develop new methods and solutions,the contributions of this thesis include the following:(1)Research and design a few-sample intent recognition model,propose the A-SNR distance metric method to build a few-sample model,improve the expressive ability of category semantics,build a multi-loss similarity function to allow the model to optimize model parameters from multiple levels,and finally pass the few-sample Fine-Tune model training method improves generalization on the target task.(2)Build a few-sample task-based dialogue design system.The dialogue design module can customize data such as intent slots and training corpus,use dialogue story scripts to achieve dialogue management,and provide model service deployment and dialogue online testing functions.The data closed-loop module can be continuously improved Dialogue data quality to build smarter dialogue systems.(3)Build an epidemic prevention and control robot based on a few-sample dialogue design system,realize multiple rounds of dialogue functions through custom training corpora and dialogue scripts,and support offline training and online deployment.After passing the visual online test,services can be released to deploy epidemic prevention and control robot.This system designs a few-sample intent recognition model to solve the problem of dependence on massive data,realizes custom businessoriented dialogue management through dialogue scripts,and realizes effective annotation and feedback of user dialogue history data based on data closed-loop modules,and mines valuable dialogue scenes and corpus.,to continuously improve the dialogue experience.