Power Grid Customer Service Intelligent System Based On Deep Learning

With the increasing amount of information held by the customer service system of the power grid,the way of dealing data is becoming increasingly unable to satisfy the needs of big data,wide business types and deep professional level.On the contrast,power grid customer service intelligent system which based on deep learning can automatically recognize entity of the work order,understand the user’s intention through semantic analysis,answer the user’s questions and recognize the user’s emotion through video,voice or text.Under the background of rapid development of artificial intelligence technology,therefore,it is necessary to conduct research on its application in existing power grid customer service systems.The research work of this paper is to develop a power grid customer service system with existing AI technology,which mainly includes three parts.Firstly,in order to solve the problem of less entity data and sparse feature distribution in the field of power grid customer service,an entity recognition module based on Bi-directional Long Short Term Memory(BiLSTM)and Conditional random field(CRF)is designed.After initially marking a small part of business entity data,a Bert+BiLSTM+CRF entity recognition model was constructed to train on it while predicting unlabeled data.Subsequently,the model was manually annotated by searching for samples that were difficult to distinguish in the unlabeled data,and the easily distinguishable samples were manually validated.These samples were then added to the training set for training,and the accuracy of entity recognition was improved by expanding the samples and retraining the difficult samples.Next,in order to give consideration to the time cost and accuracy of classification and question and answer,a text classification module based on Bert+Multilayer perceptron(MLP)and a question and answer module based on Sentence Bert and Interaction-Attitude-Feed Forward-Network(IAFN)are constructed.The user’s text will be divided into categories based on the Bert+MLP classification model.When classifing into question categories,the Sentence-Bert pre-ranking model will find several standard questions which are similar with the user’s questions.Afterwards the module uses the IAFN ranking model which based on the attention mechanism to match the most similar standard questions,returning this questions and some similar standard questions to the customer.Lastly,in order to fully utilize the video and audio information in emotion recognition,a multimodal emotion recognition module which based on Cross-Mix-Attention-Feed Forward-Network(CMFN)was designed.The user’s conversation video is first extracted from the audio track,which is transcribed by Automatic Speech Recognition(ASR)to generate text.Then,the user’s emotions during the conversation are obtained through a network that extracts the feature vectors of video,text,and sound,as well as a multimodal model designed in this paper based on the dual layer cross attention CMFN feature fusion network.The above modules were evaluated on the enterprise’s business entity data,text classification data,question and answer data,and multimodal emotion dataset.The experimental results showed that compared to traditional methods and other deep learning models in enterprises,these methods have better performance while better matching with actual business.