Research On Natural Language Understanding Method For Power Marketing Knowledge Graph

With the continuous development of natural language processing technology,intelligent dialogue systems have been widely used in various application fields.From e-commerce customer service chatbots to scenic spot voice assistants,they provide solutions and services quickly through natural language interaction with users.In recent years,the integration of information technology and intelligence in the electric power industry has become a trend,and providing retrieval services for users through intelligent question and answer has also become a trend.Natural language understanding is a core module in the dialogue system.It parses and understands natural language input from users to help the dialogue system fully understand the user’s intent and accurately extract key information,in order to provide high-quality data query and analysis services.This paper is aimed at the application of index question and answer in the field of electric power marketing.A business knowledge graph containing domain background knowledge and various index data is designed and constructed.For the index question and answer task,a natural language understanding task is designed using the graph knowledge,and a natural language understanding model that integrates knowledge encoding is implemented to identify the corresponding business domain,query intent,and slot in the user’s question,and a lightweight deployment method for the model is studied and implemented according to application requirements.The contributions of this paper mainly include three aspects.Firstly,a domain knowledge graph for the electric power marketing field is designed and constructed.Based on the entity-relationship model(ER model)of the business system database of electric power marketing,a conceptual ontology is constructed by using the relationship between the basic business data tables combed by domain experts.Then,the data tables of the business database are traversed,and the knowledge graph instance is obtained through data cleaning,data filtering,and feature selection operations,and stored in the graph database Neo4 j.The information of the knowledge graph is used to assist the natural language understanding task to set the corresponding domain,intent,and slot,and also provides additional knowledge representation for the natural language understanding model.Secondly,a few-shot natural language understanding model that integrates domain knowledge is proposed.The performance of natural language models depends on a large amount of annotated data.However,domain applications usually lack historical data,and the cost of collecting annotated data manually is too high.Therefore,this paper proposes a few-shot natural language understanding model based on encoding fusion,which enriches the representation of natural language by introducing domain knowledge of electric power marketing.Specifically,in the encoding phase,the knowledge corresponding to domain terms is fused into the model encoding to enhance the understanding of the model for unmarked identifiers in the sample.Practice shows that introducing domain knowledge can indeed make up for the lack of training samples and significantly improve the accuracy of the model.Finally,a lightweight natural language understanding model based on early exit mechanism and knowledge distillation is implemented.This paper combines the early exit mechanism and knowledge distillation to compress and optimize the natural language understanding model to ensure that it has fast inference capability in practical application scenarios while meeting the requirements of deployment cost control.The method effectively reduces the complexity and computational resource requirements of the model while keeping the performance degradation within a reasonable range.In summary,this paper designs and constructs a knowledge graph for the electric power marketing field based on electric power domain data.On this basis,a few-shot natural language understanding model based on encoding fusion is constructed.To improve the model inference speed while maintaining the model performance,this paper proposes a lightweight model construction method based on early exit mechanism and knowledge distillation.Finally,the lightweight model is integrated into the existing electric power marketing system through an API interface,which improves the efficiency of the system and the user experience of the customers.