Research On Text Sentiment Classification Based On Deep Learning

With the development of information technology and the popularity of various serviceoriented APPs,the number of comment texts has also shown a rapid growth trend.However,these comment texts contain important information and potential value.Therefore,it is necessary to excavate the potential emotional information in these comment texts through sentiment classification technology.The emotional classification studied in this paper is to mine and analyze the emotional information contained in the critical texts and then realize the correct classification of these comment texts.The purpose of this study is to accurately mine the real emotional tendencies expressed by the subjective objects to realize the value contained in these comment texts.However,the existing emotion classification methods are not perfect.In view of the problems existing in the existing traditional emotion classification models in emotion classification tasks,this paper puts forward new methods and new research ideas.The main research contents of this paper are as follows :(1)In order to solve the problems of low classification accuracy,poor emotion word capture ability and single feature extraction process of traditional emotion classification models.In this paper,we propose an EBAP model-based sentiment classification method,which uses the idea of serial feature extraction to enhance semantic understanding and knowledge representation by first generating word vectors using the pre-training model ERNIE,and then feeding the generated corresponding word vectors into a two-way gated neural network for two-way semantic feature extraction,based on which an attention mechanism is introduced to assign different weight values to the sentiment information with different degrees of importance in the comment text.The corresponding word vectors are then fed into a two-way gated neural network for two-way semantic feature extraction.The pooling layer of the convolutional neural network is then used to re-extract the most useful features and remove the redundant and useless information to reduce the factors that affect the classification results.Finally,the extracted feature vectors are fed into the Softmax classifier to obtain different sentiment polarities.The EBAP model achieves good experimental results on the Chinese sentiment classification dataset Chn Senti Corp in terms of evaluation metrics: accuracy,recall and F1,and achieves an accuracy rate of 95.33% on the test set.(2)In order to solve the problems of semantic sparsity of short text,the traditional sentiment classification model cannot fully learn to utilize the sentiment and semantic information embedded in the original comment text,and the increase in computational cost due to the increase in model complexity.In this paper,we first improve the DPCNN neural network,and then conduct step-by-step research to explore the methods of improving the DPCNN network with other neural networks and the fusion of pre-trained models,and finally propose a sentiment classification method based on the EIDBS model,which uses the ideas of multi-channel parallel extraction of features and feature fusion to first represent the pre-trained model ERNIE with knowledge-enhanced semantics The generated word vectors are then fed into the BILSTM channel and the improved DPCNN channel for feature extraction,and the extracted feature vectors from the two channels are fused with the semantic vectors for feature fusion,and finally the fused feature vectors are fed into the Softmax classifier for classification.And good experimental results were achieved on the Chinese public sentiment classification dataset Chn Senti Corp,with an accuracy of 93.92%,for the evaluation indexes such as accuracy,recall and F1,which all met the expectations.