Application Prediction And Evaluation Of Voice Robot In Intelligent Collection

With the introduction of artificial intelligence in 2019,the state has paid more and more attention to it,taking AI as an economic growth point,and using AI to improve people's livelihood.With the continuous innovation and innovation of AI technology,some chatbots have gradually entered the public's field of vision,and its appearance will change the operation mode of existing call center service.Intelligent chatbots is an inter-generational product.It relies on big data cloud computing intelligent voice as the background of research and development.It adopts AI technology such as speech recognition,NLP,dialogue management,TTS and other AI technology to achieve accurate recognition of user intentions.Fully understand the user's intention and conduct multiple rounds of dialogue with the user to enhance the loanee's willingness to repay,achieving higher execution and lower cost than manual collection.Research based on intelligent collection of chatbots has important research value in AI applications,helping credit companies to reduce costs and reduce collection complaints.Based on the research on the development status of intelligent chatbots and traditional collection models,this paper focuses on the use of collection corpus and sentiment analysis to extract customer repayment tendency.First,referring to the traditional collection scorecard model,select the relevant structural variables before,during and after the loan,and then establish the corresponding corpus according to the collection of actual application,to prepare for the subsequent withdrawal of customer repayment tendency;secondly,establish the multi-category model to extract the repayment tendency which is based on the corpus and the collection sub-category standards.Finally,the customer repayment classification effect is verified based on the structural variables and repayment tendency results.This paper aims to improve the traditional collection and scoring model based on the above research.The main research contents of this paper include the following parts:(1)This paper refers to the traditional collection model to choose variables and select the loanee's overdue behavior of pre-loan,mid-loan,post-loan and other structured data in a certain period of time from a financial credit enterprise.And then select the effective dialogue between chatbots and loanees as the research object on the basis of the traditional collection model.(2)In this paper,the influence of the variables of the traditional collection model on the collection is studied.Based on the selected variables of the traditional collection model,the collection corpus is established by semi-automatic manual annotation to define the main type and subtype classification of the customer's repayment intention.Then the vector space model is established according to the subtype of the customer's repayment intention,and the loanee's repayment tendency is extracted according to the sentiment analysis.That is,the customer's preference is manually labeled for part of customers first,and then the SVM model is used to predict the tendency of all customers.As a result,the variable dimensions of the traditional collection model are then enriched to prepare for the next step in predicting the customer's repayment results.(3)This paper combines the structural variables of the traditional collection model with the unstructured variable results obtained by sentiment analysis,and combines the Light GBM classification prediction model with sentiment analysis,and adjusts the optimal parameters of the Light GBM model through Grid Search CV method to predict the customer's repayment possibility.So that in the next case allocation,the customer with high repayment probability is assigned to the robot collection,and the customer with low repayment probability is assigned to the artificial collection.At the same time,in order to compare the classification accuracy and operation efficiency of Light GBM model,this paper also uses logistic regression,SVM and random forest to use the same data set for training,and compare the evaluation indicators and operational efficiency of each classification model.Then we found that the Light GBM model is the best model and it has the highest accuracy and the fastest speed,which can provide reasonable and effective data support for the distribution of financial credit collection cases.The research in this paper takes the financial credit collection field as the starting point,the credit record and behavioral performance actually generated by the loanee as the research object,and the dialogue between the intelligent chatbots and the customer as the carrier,gets the huge unstructured text data formed by ASR(Automatic Speech Recognition)technology,which can be transformed into a structured result that can be well analyzed through sentiment analysis.It is no longer difficult to combine structured data with unstructured data through technical means,and it makes it real value.The classification prediction model formed by the research in this paper not only makes the financial credit enterprise resources more effective allocation,but also greatly reduces the operation cost of the enterprise and reduces the complaint rate of the collection industry.This paper takes the lead in adopting multiple forecasting models in the field of intelligent collection,and extend the analysis dimension of traditional structured data with combining the unstructured data such as phonetic text,which helps to control the risk of financial credit enterprises,and has application value and reference significance for the development of intelligent collection.The modeling ideas and research methods of this paper are also of application value to other collection companies or other industries related to intelligent chatbots.