Research On Text Sentiment Analysis Based On Hybrid Neural Network

With the popularity of social media,people increasingly use the Internet to express their emotions and opinions in their daily lives,and such sentiment information can help enterprises,governments or organizations better understand and respond to public sentiment needs and feedback,as well as monitor and manage public opinion.To alleviate the problem of low attention to some sentiment words and difficulty in capturing long-distance dependencies between sentences in text sentiment analysis tasks,this paper focuses on the following aspects of work and research:1.dataset collection and organization.Based on the open-source sentiment analysis dataset as the basis,a distributed and highly available crawler program is developed and implemented to extract and optimize the review texts of Taobao,VWAP and takeaway platforms,and a total of 22000 Chinese sentiment datasets are obtained.Also,40258 Chinese sentiment dictionaries containing sentiment polarity were collected and obtained.2.A DBGA sentiment analysis model based on BERT and gated attention optimization is proposed to solve the problems of traditional neural network-based sentiment analysis methods that are difficult to capture the long-distance dependencies between sentences and the low attention of some sentiment words.The model utilizes the bidirectional representation of BERT to obtain richer semantic features,combines a bidirectional gated memory network layer and a self-attention mechanism layer,and proposes a subword selection algorithm(BDSS)for fused sentiment words compatible with the BERT structure.The experimental results show that the accuracy of this model is improved by 2.07%over the baseline BERT model.3.propose a DXMR sentiment analysis model based on XLNet with hybrid network optimization.To further improve the accuracy of sentiment analysis,the DXMR optimization model is proposed.The XLNet model,which is superior to BERT,is utilized for vectorization,and then a hybrid neural network(gated attention and multi-channel text convolutional network with proven effectiveness of DBGA)is used to obtain richer text features,which further improves the accuracy and generalization ability of sentiment analysis.Also,an XLNet-based subword selection algorithm(XDSS)for fused sentiment words is proposed to make the network model more focused on sentiment words,which can assist DXMR to improve the performance of the model.Through the validation of comparison and ablation experiments,the accuracy of the DXMR optimized model is improved by 3.67% over the baseline and also 1.6% over the DBGA model,which illustrates its effectiveness and superiority.4.Design and implementation of the prototype system for sentiment analysis.In order to apply the methods proposed in this paper to real life,the DBGA optimization model proposed in Chapter 4 and the DXMR optimization model proposed in Chapter 5 are used as the core of the sentiment analysis technique,the optimal model is loaded and encapsulated into an interface.Finally,a comparison of the sentiment analysis results of each model is shown,and the feasibility of the proposed model is demonstrated.