Research And Application Of Sentence Similarity Calculation Method Based On Convolutional Neural Network

Sentence similarity calculation is the basis of many natural language processing tasks,such as machine translation,plagiarism detection,query sorting and question answering.The accuracy of sentence similarity calculation directly affects the performance of related systems.Therefore,how to improve the accuracy of sentence similarity calculation becomes an urgent problem to be solved.Traditional natural language processing methods mainly use the shallow features specified by human.And the method based on the neural network model can learn features from large-scale texts without any feature of artificial construction,In this way,the deep semantic information of the sentence can be obtained.Because convolutional neural network has many problems such as small receptive field and single convolution in feature extraction.This paper proposes two models to improve the above problems and stud ies the application of sentence similarity calculation in knowledge q&a,it mainly studies the following three aspects:(1)A sentence similarity computing model combining convolution and expansion convolution named CNN-IDCNN is proposed.(convolutional neural networks-iterated dilated convolutional neural networks).for convolution neural network at the time of extracting text feature between receptive field is narrow,levels of pooling method exists the shortage of the information loss,put forward on the basis of ordinary convolution module to join expansion convolution module method to extract the sentence semantic information over long distances.Let the characteristic expressions of sentences contain word,short n-gram,long n-gram information,enriching sentence feature representation from multiple granularity to improve the computational performance of similarity.(2)We improved our model in convolutional neural network based on syntax tree convolutional neural networks.convolutional neural network is to extract features by convolving adjacent words through a window of specific size,because the convolution window is usually small and just convolution of adjacent words.And this approach is not flexible enough,it ignores the combination of words in the original sentence which are not adjacent to each other but have strong semantic relevance,for this purpose,the convolutional neural network integrated with syntax tree is used in this chapter,the above problem is improved by applying the convolution kernel to the word combination obtained from the dependency syntax tree.Experimental results show that the proposed method is superior to the traditional method,and better than some methods based on convolutional neural network model proposed in recent years.(3)We implemented the question answering system based on question matching technology.This paper mainly studies the application of similarity calculation in question answering system,compare similarities between user question and the candidate sentences through sentence similarity models which proposed in this paper.we calculate the candidate sentence that most resembles the semantics of the sentence asked by the user and then correspond the answer to the user.