Research On Text Sentiment Analysis Method Based On Attention Mechanism And Neural Network

With the continuous development of science and technology,more and more users share their opinions online.A large amount of user emotional information is hidden on social networking sites and shopping sites.How to dig out effective information from these data and make use of it has become a research hotspot.Text sentiment analysis is a process of exploring and analyzing the deep-level sentiment information in the text,and find and use the effective information in the text in a certain way.However,the existing sentiment analysis models cannot effectively extract the deep-level sentiment information of the text,and lack the ability to extract important information in the text.In order to further improve the level of sentiment analysis,according to the characteristics of different models,this article proposes corresponding improvement methods,The specific research content is as follows:(1)In view of the problems that traditional recurrent neural networks may cause information loss and gradient dispersion when processing long sequences,and the current deep learning-based analysis model does not strengthen the attention to emotional words,this paper designs the BiAGRU model,which combines the attention mechanism and the two-way The recurrent neural network is combined,and the attention mechanism score is used to replace the value of the update gate in the GRU model,and the two-way propagation is carried out.Give more attention to the more important words in the text,enhance the relative weight of important words,and strengthen the learning of long-sequence semantic information.Experimental results show that compared with some traditional methods,the algorithm has improved in many aspects.(2)When using the ReLU activation function for the convolutional neural network,if the gradient value is negative,neuron necrosis may occur.This article uses the PReLU activation function to replace the ReLU activation function;only keep for the maximum pooling layer A maximum value,ignoring other strong features,will lose a lot of feature information.This article uses K-Max Pooling to improve the pooling layer.By retaining multiple strong features,part of the text information is retained.These improvements make the text volume Product neural networks have better text feature extraction capabilities and are more suitable for text feature extraction.(3)Aiming at the feature that convolutional neural network ignores contextual semantic information,but has strong sensitivity to local information,while bidirectional recurrent neural network can extract global information well,this paper designs a combination of NTexCNN model and BiAGRU model Multi-channel text sentiment analysis model TC-FFA-BiAGRU model.First,use the Jieba word segmentation tool and the GloVe word vector model to preprocess the text data;secondly,use the improved NTexCNN model and BiAGRU model to extract the local information and global information of the input vector;then use the feature fusion layer to combine the two features The vectors are fused,and the fused features are sent to the force-forward attention mechanism layer,and finally the emotion classification is realized through the fully connected layer.Experiments show that compared to traditional network models and some patchwork models such as CNN-BGRU and CNN-BiGRU,the model in this article has a better sentiment classification effect.