Research And Application Of Semantic Matching Based On Deep Learning In FAQ Question Answering

FAQ(Frequently Asked Questions)is an important component of an intelligent customer service system that answers Frequently Asked Questions in the business domain and improves service efficiency while saving labor costs.FAQ is a retrieval question-answering system for frequently asked questions.The retrieval process is divided into two steps:initial recall and precise ranking,and the key technology of which is semantic matching.With the expansion of customer service and the improvement of intelligence requirements,the current semantic matching algorithms have many limitations:Firstly,the model has poor adaptability among different fields,but the corpus data that can be used to train the special model is limited.Secondly,the traditional semantic matching algorithm used in the recall phase is weak in the deep semantic matching and will miss some positive samples,which directly affects the matching space in the ranking phase.Finally,in the ranking stage,the deep semantic matching algorithm does not capture the "fine-grained" features sufficiently,and the matching accuracy is not high in the FAQ scene,and there is deviation in the understanding of user intent.Therefore,this thesis improves and optimizes the models in the recall phase and the sorting phase,aiming at the migration adaptability and recognition accuracy of the key semantic matching technology in FAQ question answering system.Specific contents include:(1)In order to solve the problem of low recall rate in FAQ scenario due to the lack of semantic matching algorithm in recall stage,a hybrid recall algorithm is designed in this thesis.First,Siamese-Bert,a deep semantic vector encoder trained by Siamese network structure,is introduced to capture deep semantics.Then,it is combined with traditional semantic matching algorithm BM25 to exploit the common advantages of both algorithms.By using the question data sampled from the real FAQ data of the enterprise,the hybrid recall algorithm can improve the recall rate compared with the single use of BM25 algorithm and deep semantic vector matching algorithm,and has a good performance in the data matching in different fields.(2)In order to solve the problems of insufficient capture of "fine-grained" features and low matching accuracy in the ranking stage,an improved multi-feature fusion deep semantic matching algorithm MFF-BERT based on BERT model is proposed in this thesis.Firstly,the model was made to pay attention to the "fine-grained" features by adding artificial monitoring signals,and then the "category" recognition features were added in combination with the FAQ scene matching features.Finally,the two features were fused with the original BERT model.The experimental results show that,in the FAQ scenario,the accuracy of the improved multi-feature fusion deep semantic matching algorithm in this thesis is 5.3%higher than that of the original BERT model,and the algorithm has a good improvement effect in different fields of data.(3)Based on the algorithm improvement of the above two points,this thesis designs and implements the matching related modules of FAQ question answering system,including the knowledge base building module that provides data support,the recall module designed based on the hybrid mode recall algorithm,and the ranking module designed based on the deep semantic matching algorithm of multi-feature fusion.The practical value of the algorithm is verified in the actual scene,which provides a new solution for the improvement and application of the matching algorithm in FAQ scene.