| Voice is the most natural human-computer interaction,the most suitable user interaction.Amazon echo and Google home smart speakers can control home appliances by voice.The water heater embedded in the question-answering system is more intelligent,bringing great convenience and security to users.In recent years,the rapid development of natural language processing technology,intelligent question answering system is the natural language processing difficulties and hot spots,the integrated use of natural language understanding,question matching,information retrieval technology to understand user intent and quickly and accurately to answer.This dissertation mainly studies the key technologies in question answering system such as topic classification,sentence similarity calculation.There are three main difficulties in the research of intelligent question answering system for electric water heater:(1)The user’s intention to identify,that is,the subject of the sentence classification.The interaction between the user and the electric water heater use more short sentences,omitting more leads to the lack of thematic features,the traditional machine learning algorithm is suitable for long sentence classification,can not effectively extract thematic features of the short sentence,resulting in poor classification.(2)The sample imbalance of corpus concentrated in the electric water heater leads to the poor classification accuracy of the sub-sample.(3)Answer matching.Due to the sparseness of the user’s sentence matrices,and the traditional sentence similarity calculation method ignores the dependency between words and words and the deep semantics of sentences,the calculation error of statement similarity is large,which leads to a high response matching error rate.In this dissertation,intelligent water heater QA system for more than three targeted research,the main contributions and innovations include the following:(1)In this dissertation,the convolution neural network is applied to the topic classification of user statements,and 98.42%of the topic classification accuracy is obtained on the water heater corpus.(2)In order to solve the problem of imbalanced samples in the corpus of electric water heaters,the SMOTE algorithm is applied to sample corpus and balance the corpus,which improves the accuracy of the topic classification to 99.25%.(3)In this dissertation,word2vec,a language model of deep learning,is applied to the similarity calculation of sentences.User’s sentence similarity algorithm based on word2vec and part-of-speech weights is proposed according to the topic features to improve the accuracy of statement similarity calculation.In this dissertation,topic classification accuracy rate of 99.25%based on CNN compare with the traditional method to improve 11.58%,user sentence similarity calculation is 10.84%higher than the traditional method accuracy.At the end of this paper,the limitations of the existing methods and the prospects of the next generation intelligent water heater QA system are analyzed. |
