Sentiment Analysis Of Text Reviews Based On Deep Learning

Sentiment analysis is a kind of text classification,which can help people explore the theory and application of emotional intelligence in product marketing,public opinion monitoring,social network analysis and other aspects by determining the emotional tendency of positive emotion,negative emotion and other aspects of text.When performing sentiment analysis tasks,it is important to design models that meet the characteristics of the data and the requirements of the task.Multi-strategy combination methods such as model combination,model integration,and introduction of prior knowledge are commonly used improvement methods for text classification models.Based on the research of previous scholars,it is proved that model fusion is not only feasible,but also very effective.Therefore,this thesis mainly studies and discusses the model fusion method with RNN algorithm as the main model,and verifies it through experiments,hoping to provide a reasonable reference and reference for further improving the accuracy and efficiency of sentiment classification.In this thesis,a comparative experiment of machine learning models is designed to demonstrate the advantages of deep learning algorithms in handling sentiment analysis tasks.In the case of changing the variable of word embedding dimension,the effect of single model sentiment classification of different network structures is compared,which provides a reference for designing the network layer order of fusion model.In order to improve the classification efficiency and accuracy of a single model and mine more comprehensive feature information,this thesis proposes two fusion models based on RNN.One is the Text RCNN model combining RNN with CNN,and the other is a Text RNN＿Att model combining RNN and Attention.The Text RCNN model uses the Bi LSTM model to extract text global features in both directions,and then access the module of CNN for feature dimensionality reduction,unlike other fusion models,Text RCNN only uses the pooled part of CNN.Text RNN＿Att model improves RNN by integrating attention mechanisms,and uses the weight calculation of attention mechanisms to filter out important contextual information from inputs.These two improved RNN-based models have higher classification effect and accuracy than single network models in text sentiment analysis experiments.In particular,the Text RCNN model has the highest classification accuracy,accuracy,recall rate and F₁score on text review data,reaching 95.16%,95.15%,95.14%and 95.15%,respectively,which are higher than the Text CRNN model,and the accuracy rate is 8.61%higher than that of RNN.The performance of model sentiment classification is improved,and the rationality of the sequential design of the network layer is also verified.This thesis also discusses the performance of static word vector representation and dynamic word vector representation models in sentiment classification.Both BERT and Glove models are techniques used to generate word vector representations,and they differ in terms of algorithm principles,context considerations,and applicable scenarios.The BERT model is essentially a pre-trained language model that integrates the Focus and Transformer encoders,and is widely used in the NLP field.Without combining other network models,the BERT model can achieve a classification accuracy of 89.80%on the comment dataset used in this thesis,and the classification accuracy is higher than that of the static representation model,indicating that the dynamic extraction model has significant advantages in dealing with text problems.