Research And Implementation Of Joint Classification And Matching Model For FAQ Question And Answer

Since the new century,the technology of information technology,especially artificial intelligence,has developed rapidly.People can access information services through smart de-vices such as mobile phones and tablet PCs anytime and anywhere.The number of enterprise customer service has been unable to meet the demand of more and more people to obtain information.The goal of the intelligent customer service system is to enable users to obtain information through natural language and computer-friendly interaction,thus reducing the pressure of manual customer service.In order to answer user's questions more accurately,companies usually organize FAQ(Frequently Asked Questions)data sets,which generally contain standard questions and answers to standard questions.By calculating the similarity between the user question and the standard question,and returning the answer corresponding to the user's closest standard question,the user's request for information is satisfied.As user questions continue to accumulate,companies often document the mapping of these historical user questions to standard questions.For FAQ data set with historical user questions,current methods fall into two main categories.The first category is to use the standard question as a label,and to construct a categorizer corresponding to the standard selection problem by constructing a sentence representation of the historical user problem.The second category is to calculate the semantic distance between historical user questions and standard questions to get the best answer.Both of these methods have been widely studied and applied,but at the same time,the following problems exist:First,the FAQ datas is generally short in length,lack of context information,and lack of grammatical structure.The current method cannot be solved well;It is the classification method that cannot use the standard question as the important information contained in the label itself,and the negative sampling of the matching method is a major problem,and it is difficult to select a negative example that really needs to be distinguished.For the first problem,this thesis designs a joint model of width neural network.The model combines a word-grained convolutional neural network with a word-granular-based bidirectional long short term memory network to construct a classifier.The model can not only capture the phrase features,grammatical features,etc.,but also learn the time series features well,which largely alleviates the problem of missing context information.Compared with other models,the joint classification model proposed in this thesis has achieved the best results in real enterprise customer service data,and the model has also been put into use by the enterprise.For the second question,this thesis designs a model for joint classification and matching.The model effectively combines the classification method and the matching method,and can not only select the negative examples that really need to be distinguished,but also can effectively use the information of the standard questions.By sharing the word vector,the historical user problem is used to construct the classifier while the sentence is being represented,and the metric calculation is also performed with the sentence representation of the standard question.Compared with many current models,the model shows the best results in both Chinese and English data sets.Based on the above two types of models,this thesis designs and implements the FAQ question and answer system.The system consists of five modules:FAQ library management module,FAQ answer calculation module,Web design module,feedback module,and log module.The main functions of the system are:user selection model interaction,user feedback,log recording,and so on.