Research On Text Sentiment Analysis Based On Feature Fusion Of TCN And Bi-LSTM

With the development of mobile technol ogy and the increase in the number of netizens,the information on the Internet is increasing day by day,and text is one of the most important manifestations of these data.Extracting and using valuable subjective information from massive text data is a very popular research topic in the field of natural language processing today.Among them,the analysis and research on the subjective sentiment implied in the text can provide useful guidance for user recommendation,public opinion analysis and other fields.Although the existing text sentiment analysis methods have achieved certain results,they still have some shortcomings: Relying on the sentiment lexicon to judge sentiment polarity,its accuracy is too dependent on the quality of the dictionary,and it is not easy to maintain;When using machine learning algorithms to extract emotional features,the structure of feature engineering is extremely complex and mostly ignores the sequence features of the text;Commonly used neural network models have the disadvantages that it is difficult to solve long-term dependence and insufficient use of contextual information.Therefore,based on the summary and analysis of these methods,this article conducts in-depth research on the construction and application of text sentiment analysis models based on deep learning.The main work and innovation are as follows:(1)Constructed the T-BiLSTM sentiment analysis model: The proposed model is mainly aimed at the defects of the existing neural network model.Apply the advantages of Temporal Convolutional Network(TCN)in processing time series problems to text processing to learn the sequence features of text,and use the advantage of bi-directional long-term short-term memory network(Bi-LSTM)in learning text context information to extract the bidirectional semantic dependence of text.Through the design idea of the residual network,the features extracted from the two network layers are merged for classification.Experiments have verified the effectiveness of the model.(2)Constructed the TCN-BA sentiment analysis model: Considering that the self-attention mechanism can capture the internal structure of sentences when processing text data,it is introduced into the T-BiLSTM model.By assigning different weights to the vectors that affect the judgment of emotional polarity,the dominant features are highlighted,thereby improving the accuracy of the model in recognizing emotional polarity.Compared with the T-BiLSTM model,the accuracy of the TCN-BA model with the introduction of the self-attention mechanism in the two data set experiments in the same experimental environment has increased by 1.54% and1.83%.(3)Sentiment analysis of Weibo texts during the epidemic: In this thesis,we analyzed the microblog texts during the epidemic.The data were crawled from microblogging platforms based on keywords related to the novel coronavirus epidemic,and the crawled data were pre-processed,labeled and analyzed to construct a dataset of microblog texts during the epidemic.After that,the two models proposed in this thesis were applied to the data set,and the experimental data proved that the accuracy and loss performance of the t wo models are better than other benchmark models.