Research On Sentence Similarity Algorithm For Intelligent Question Answering System

Intelligent question and answer system is a kind of computer program which can answer the natural language question sentences that users input,and it is an artificial intelligence service system which integrates the natural language processing,information retrieval,semantic analysis,machine learning and other technologies.Because it uses natural language interaction,it improves user experience,so it is widely used in service and consulting field.The core function of intelligent question and answer system is to match the questions raised by users with the questions in FAQ database,find the most similar items with the user input question sentences semantics,and finally return the matching questions and corresponding answers to the users.The quality of FAQ is very important for the question and answer system.On the one hand,due to the variety of user questioning methods and the continuous emergence of new words,the "question and answer" pairs in FAQ database need to be updated constantly,so as to be more accurate and more matching users' questions;on the other hand,the scale of constantly newer FAQ library will gradually increase When the boundaries between similar questions begin to blur,the answer returned to the user may appear to be answered in a non-question situation.Aiming at the deficiency of the accuracy of FAQ,this thesis studies sentence similarity algorithm from two aspects: word vector and deep learning.The main work is as follows:(1)Research on sentence similarity algorithm based on word vector.For the convenience of computer processing,Chinese words need to be converted into word vector form.The similarity of sentences can be calculated by the difference between the word vectors of two sentences.Therefore,the distance between synonyms and synonyms should be closer,and the distance between antonyms should be farther.In addition,we need to calculate the word vector of new words that are not in the vocabulary.Based on the synonym dictionary of Harbin Institute of technology,the near antonym vector is mapped to the near synonym space,which makes the near synonym vector closer and the antonym vector farther.By combining BPE algorithm with NGram method to deal with the unknown new words,the model can recognize the near antonym and new words more accurately.In this thesis,based on the traditional WMD distance algorithm,the word frequency weight and part of speech weight are combined to calculate the sentence similarity as a new weight coefficient.(2)Research on sentence similarity algorithm based on deep learning.When the scale of FAQ gradually increases,the model based on deep learning is used to calculate the sentence similarity.Firstly,based on bigru twin network model,a twin neural network model of bigru combined with attention mechanism is proposed to calculate sentence similarity.Secondly,Albert model and bigru are used to calculate sentence similarity.Finally,a multi angle fusion strategy is proposed,which not only adds the cooccurrence information of words and words as the minimum semantic unit,but also improves the performance of twin neural network.In order to improve the accuracy and stability of sentence similarity calculation,the network model and syntactic dependency model are weighted.(3)Design and implementation of intelligent question-answering system for public inspection.The similarity calculation method based on twin network has important practical value.In order to verify the effectiveness of the above method in the actual scene,this thesis designs and implements the intelligent question and answer system of public inspection service.Firstly,the FAQ Library of the system is briefly introduced.Secondly,the overall framework,function flow and similarity calculation flow of the system are designed.Finally,the system is implemented and demonstrated by a specific example of question answering matching.Different from the existing intelligent question answering system,the design style of this system is simple and easy to operate.It supports the return of the most similar questions and their answers,and better meets the needs of users.