Research And Application Of Task-oriented Dialogue System Algorithm Based On Deep Learning

With the wide application of task-oriented dialogue systems in customer service,smart homes,and other domains,the task-oriented multiturn dialogue technology,characterized by wide application scope and high service efficiency,has attracted much attention.Task-oriented dialogue systems are typically developed using a pipeline-based framework,which includes four modules:natural language understanding(NLU),dialogue state tracking(DST),policy learning,and natural language generation(NLG).Among them,the dialogue state tracking module is the core module of the pipeline-based task-oriented multi-turn dialogue system.Currently,in cross-domain few-shot scenarios,the DST module faces challenges in tracking capability.Additionally,the development framework for taskoriented multi-turn dialogue systems in new domains is still worth exploring.To address the issue of low recognition accuracy in the DST module in cross-domain few-shot scenarios,this paper proposes a Fusion Prompting Dialogue State Tracking(FP-DST)model that combines visible templates with continuously adaptive templates.The FP-DST model utilizes visible templates to guide the initialization of slot descriptions in the continuously adaptive templates.Then,the continuously adaptive templates further optimize the representation of slot descriptions,effectively leveraging prior knowledge accumulated by large pre-trained language models to achieve higher tracking accuracy in few shot scenarios.The feasibility of the proposed model is validated through experiments based on the joint accuracy metric.To meet the development requirements of task-oriented multi-turn dialogue systems in new domains,this paper designs a pipeline-based framework that supports multi-domain scenarios.The framework incorporates BERT encoders and GRU decoders with soft-gated copy mechanism in both the NLU and NLG modules,enhancing the crossdomain contextual understanding and response generation diversity in cross-domain settings.Experimental results on the Chinese multi-domain multi-turn dialogue dataset,CrossWOZ,demonstrate that the framework achieves an intent recognition accuracy of 80.72%across different domains and a BLEU similarity of 32.17%in response generation compared to the label,validating the practicality of the proposed framework.Finally,targeting the Winter Olympics-themed tourism application scenario,this paper has developed a Chinese task-oriented multi-turn dialogue prototype service specifically designed for Winter Olympicsthemed tourism.A Winter Olympics-themed dialogue corpus dataset was constructed for Beijing and Zhangjiakou,along with ontology data for hotels,attractions,restaurants.The proposed development framework was employed to develop the Winter Olympics dialogue prototype service,including training the algorithm modules,fine-tuning,and system deployment.Subsequently,functional testing was conducted on both single-domain and cross-domain test cases.The dialogue system achieved an intent recognition accuracy of 99.7%on the self-built Winter Olympics dialogue corpus dataset.