Research On Deep Percolation Network For Sentence Classificatio

Sentence classification is a fundamental task in natural language processing.Sentence classification can effectively support tasks such as sentiment classification,relation extraction,and news topic classification.Existing sentence classification models employ deep neural networks to automatically extract abstract semantic features from original sentences.In convolutional neural networks,each output corresponds to an abstract representation of the entire input.Since some words in the input have stronger classification features,it is easy to confuse the abstract representations of other words after convolution operation,thereby weakening the classification effect of words.In addition,in the deep neural network,by stacking multiple layers of neural networks,on the one hand,the high-order semantics of the sentence can be obtained;on the other hand,due to the problem of gradient disappearance,the deep structure will lead to performance degradation.In order to solve these problems,this paper designs a deep penetration network structure.This model enables salient features to be transmitted in deeper network structures,builds longer semantic dependencies,and retains useful semantic features,so as to better obtain high-order semantic features in sentences and improve the performance of sentence classification.The main work of this paper is as follows:(1)Research on deep penetration network(DPN)based on upsampling.In this paper,a deep penetration network based on upsampling is proposed to solve the problems of semantic feature vanishing caused by deep structure in neural network,difficulty in extracting high-order semantic features from network model,and degradation of model performance.In this network,the semantic features of the text are strengthened by up-sampling before the convolution operation,so that the salient features can penetrate layer by layer in the deep structure,and the semantic dependence between salient features is constructed in the high-level network.Infiltration network can effectively extract high order semantic information in sentences and solve the semantic loss caused by deep neural network in sentence classification.Experimental results show that the semantic feature retention ability and semantic dependency construction ability of deep penetration network are significantly improved compared with traditional methods.(2)Sentence classification model based on deep penetration network.Based on the deep penetration network,we design and implement a deep neural network model with two channels.Based on the idea of pyramid model,this model can effectively utilize features of different depths in the network and learn semantic features of different granularity in sentences to support the task of sentence classification.The performance of this model is better than other models in many tasks,such as news text classification,emotion classification and relationship recognition.The experimental results show that the deep penetration network has stable performance with the deepening of the structure.It has achieved better performance in sentence classification and better results in related work.