Multi-channel Text Matching Approaches Based On Deep Neural Network

In the 21 st century,the application of natural language processing has been expanding in the field of computer science.The demand of people in the field of artificial intelligence such as information retrieval,question answering systems,dialog systems,intelligent customer service,and machine translation start to appear.These tasks can be largely abstracted into text matching problems.Hence,text matching technology came into being.Text matching is a key technology in the field of natural language processing and can be applied to a large number of natural language processing tasks.In recent years,computer vision technology has developed rapidly,and it has gradually become a new research direction to introduce computer vision technology into text matching tasks.By studying the classical deep text matching model in recent years,this thesis finds that the text feature extraction does not consider the local information such as word,word,and phrase as well as the temporal information of the text.It is also found that in the fusion and extraction of text matching or text interaction information,there is no deeper extraction of its features,and no better use of local information of text matching.Firstly,aiming at text feature extraction,this thesis designs a deep text feature extraction method based on convolutional neural network.The newly designed method can simultaneously extract text word features,local features such as words,phrases,or text timing features.Secondly,aiming at the problem of fusion and extraction of text matching information or text interaction information,this thesis designs a method of multi-channel feature stacking and extraction.Based on text matching calculation and text interaction calculation,this thesis constructs the multiple perspectives based on convolutional neural network channel text matching model D-MPCPM and attention under the mechanism of multimodal interaction multiple points of view A-MPCIM text matching model,among them,the D-MPCPM model mainly through different matching method or model parameters matching matrix,multiple points of view and deep convolution neural network is adopted to improve the matching feature extraction;The A-MPCIM model is mainly about the interaction between two text features.The interaction method is based on the matching matrix to calculate the Attention weight.After the interaction,the two text features are fused with different methods to obtain interactive features from multiple perspectives.Finally,the newly designed models are validated on two open data sets.A comparison experiment is conducted for deep text feature extraction methods.The experimental results show that the deep text feature extraction method based on convolutional neural network is more stable.It has stronger generalization ability,and can extract more text features.Then,a comparative experiment of multi-channel feature stacking and extraction is conducted.The experimental results show that multi-channel feature stacking and extraction are superior to Bi LSTM extraction.Moreover,a comparative experiment of the deep text matching model is carried out.The experimental results show that the D-MPCPM model and A-MPCIM model constructed in this thesis are superior to the typical deep text matching models in recent years.