Method And Application Of Text Sentiment Analysis Based On Fusion Of Surface And Deep Features

With the development of microblog and other network platforms,more and more people are willing to express their feelings and opinions,Therefore,a large number of subjective texts are accumulated in the Internet.By mining the information in the text,we can timely understand people's emotional tendency to a certain problem.Text sentiment analysis is to automatically identify the subjective views in the document and judge the emotional tendency of the whole document.In recent years,the best performance of text sentiment analysis method is the deep neural network.Especially the Convolutional Neural Networks and Bi-directional Long Short Term Memory.The core idea of these methods is to extract deep semantic local features and context global features from text data for sentiment classification.However,there are also some shortcomings,such as not considering the different effects of different features on classification accuracy,and the emotional information contained in words with different parts of speech.Aiming at the problems of deep neural network model of affective analysis,this paper studies the model and algorithm of affective analysis.The main research contents are as follows:(1)In order to solve the problem that the feature extraction of the model based on CNN and Bi LSTM features is not comprehensive and the influence of each feature is not considered,a fusion of Local and Global Key Features is proposed.This method extracts more comprehensive local feature representation by increasing the types of convolution kernels,and introduces attention mechanism to distinguish the different effects of feature representation on classification results.Experimental results on three datasets show that the accuracy of the proposed method is about 4% higher than that of the original method.(2)Aiming at the problem that the fusion of Local and Global Key Features does not consider the word vector dimension information of words,an improved fusion of Variable Convolution and Bi LSTM Key Features method is proposed.This method changes the convolution mode of convolution layer,which is vertically convoluted from the sentence direction of the document and horizontal convolution in the dimension direction of the word vector;In the pooling layer,the important information in the feature representation is retained by using the maximum pooling and average pooling.Experimental results show that the classification accuracy of this method is improved by about 1% compared with LGKF model.(3)In the fusion of Variable Convolution and Bi LSTM Key Features method,the network model does not consider the surface features of documents,and proposes a method to Merging Surface and Deep Features of documents.The method adds part of speech vector to the input data,which makes the emotion analysis model extract more accurate deep semantic feature representation from the word vector and part of speech vector.Experimental results on two English datasets show that the classification accuracy of this method is higher than other models.