Research Of Textual Similarity Algorithm Based On Siamese Recurrent Neural Network And Convolutional Neural Network

A large amount of information are texts in today's life.The textual similarity algorithm can extract meaning and potential information from these texts,and apply the information to actual systems,such as search engine,automatic textual summarization,automatic translation.The research of textual similarity algorithm provides technical support for the applications of text-based.Traditional text similarity algorithms are easy to implement,but these algorithms are challenging to extract local and global features of the text.The research of textual similarity algorithm adopted the deep learning with the development of deep learning after 2012 gradually.However,the current textual similarity algorithm based on deep learning is inadequate for global and local textual features extraction,and there is no distinction between the effectiveness of global and local textual features extraction.This paper proposes a Siamese Recurrent Neural Network and Convolutional Neural Network to solve these shortcomings.The Siamese Recurrent Neural Network(SRNN)model improves to be suitable for the extraction of global textual features,such as: removing the Embedding layer,using a decreasing multilayer Bi-directional LSTM.This paper proposes a Multichannel Input Siamese Text Convolutional Neural Network(MSCNN)model,which used to study the relationship between sequential and reverse text.The proposed model is used to extract local textual features mainly.This paper performs weight learning on the SRNN and the MSCNN to study which of the two models is more useful for the textual similarity task.In order to verify the model,this paper uses the BERT pre-training model and experiments on the LCQMC data set.The experimental results show that the model of this paper has a higher accuracy than the state-of-the-art model.Finally,using the model to a content-based recommendation task and giving the recommended results by the model.From the recommended results,the model of this paper can find more similar texts.