Text Sentiment Analysis Based On Feature Fusion

Internet technology advances at a rapid pace to promote the rapid development of the Internet field such as social networking,and particularly important is the increasingly perfect mobile Internet technology,which allows users to easily express their opinions online and publish interesting things around them,and these are mostly texts.Structured information,sentiment analysis of text information can help organizations such as enterprises to understand user preferences,adjust the direction of development to adapt to the market.This article focuses on how to improve the classification effect of text sentiment analysis,from document-level text sentiment analysis and aspect-level text sentiment analysis.The main contents are as follows:On the one hand,document-level text sentiment analysis to build'a hybrid neural network layer-using the attention mechanism based Capsule Network(Capsule Network,CapNets)model and Bi-directional Gate Recurrent Unit(BiGRU)model to fully extract local features Information and global feature information,adaptively perceive context information and extract feature information that affects text sentiment analysis,and fuse feature information learned by the model.Then,the text information represented by two different word vectors(Word Vector)respectively passed through the mixed neural network layer,and the results were further fused,and the fusion results were classified with Softmax.This model is compared with the benchm ark model.The running results of the model show that the proposed model is feasible.On the other hand,aspect-level text sentiment analysis considers the in fluence of aspect item position information.If words in aspect items appear in the text information,the adjacent words are given higher weight,and aspect item position information is embedded as part of the input,Use a hybrid neural network to extract higher-level semantic feature representations and perform feature fusion.On the other hand,each word in the aspect item plays a different role in judging the sentiment tendency.Therefore,BiGRU is used to extract the information embedded in the aspect item and calculate the attention weight represented by the text information corresponding to each word in the aspect item.The effectiveness is proved on the SemEval-2014 Task4 standard data set.