Research And Application Of Knowledge Point Matching Algorithm For The Dialogue Scene Of Government Customer Service

The government service hotline is set up for the masses.It is generally used for consulting on administrative functions,responsibilities,policies and regulations,understanding or soliciting opinions and suggestions on administrative management,social management,public services and seeking legal aid.Due to the different problems the masses calling the hotline,the time needed to solve them varies.It is of great significance to use deep learning algorithm to improve the operator’s problem-solving efficiency and provide more convenient and efficient services.Based on the analysis of the real data of Nanjing 12345 hotline,this thesis constructs a knowledge point matching model based on deep learning.In the process of customer service and user call,it automatically matches the relevant knowledge points and recommend them to customer service,so that customer service personnel can quickly answer the user’s questions,and improve the efficiency of government service hotline.Firstly,this thesis analyzes the process of knowledge point matching in the scene of government dialogue,and divides the task into data collection,building ASR model,building text error correction model,text data analysis,building text matching model,knowledge distillation and other modules.Through the cooperation with Nanjing municipal government,we obtained 12345 hotline data of Nanjing.The speech data is preprocessed and feature extracted,and the acoustic model and language model are constructed to complete automatic speech recognition,and the conversation speech is transcribed into text.Because there are homophones and other errors in speech transcripts,a text error correction model based on bilstm-crf is constructed to filter the transcripts.Finally,the filtered text is processed and analyzed.A knowledge point matching model based on Siamese Bert and Triple Loss is proposed.By analyzing the task of knowledge point matching in government dialogue scene,this thesis adopts siamese structure and introduces the triple loss function.And uses the pre-training language model Bert with strong text coding ability.Compared with the common text matching models ESIM and ARC-I,the recall rate is increased by about 8%,and compared with the text matching model fine tuned by Bert,the recall rate is increased by about 5%.Through further experiments on different pre-training language models,it is found that using Ro BERTa-wwm-ext can improve the recall rate to about 77%.Finally,through the model reasoning time comparison experiment,it is found that the single text coding time of the pre training model is up to 40 ms,while the single text coding time of ESIM and arc-i model is only about 8 ms.A multi-stage knowledge distillation method for customer service dialogue scenario is proposed.Combined with the matching characteristics of knowledge points in government dialogue scene,knowledge distillation is divided into two stages: semantic distillation stage of high-frequency knowledge points and fine-tuning stage of students’ network.High frequency knowledge point classification task is used to complete the semantic distillation stage,and knowledge point matching task is used to build the student network fine-tuning stage.Through the comparative experiment,it is found that the recall rate of student network after distillation is about 4% lower than that of teacher network,but the parameter quantity is only half of that of teacher network,and the reasoning time is twice as fast as that of teacher network,reaching 20 ms.