Research And Application Of Sentiment Analysis Technology Based On Deep Learning

With the rapid development of the Internet,people increasingly like to express their emotions and opinions on various e-commerce platforms and social media platforms.Internet platforms like Weibo,Facebook,Taobao,e Bay,etc.generate massive text comments every day.The big data generated by the Internet has great potential.If sentiment analysis technology can be used to obtain the sentiment tendency of the publisher in the text,it will create great value for the society.With the rapid development of contemporary hardware performance,deep learning technology has ushered in the best era.Sentiment analysis technology based on deep learning has excellent performance in prediction accuracy and efficiency,and it has received more and more attention from research scholars.In order to further explore sentiment analysis technology based on deep learning,this thesis focuses on the research on coarse-grained sentiment analysis tasks and aspect-level sentiment analysis tasks.The main work includes the following points:(1)For coarse-grained sentiment analysis tasks,this thesis proposes an improved capsule network model.The capsule network model cannot effectively extract the contextual semantic information of long texts,and cannot pay attention to the key content in the text,so this thesis proposes an embedding enhancement module.In this module,the BLSTM layer uses two LSTM models to cyclically input text from both positive and negative directions,which can effectively model the text context features.The attention layer can calculate the importance of each word,which helps reduce the influence of irrelevant words on the global characteristics of the text.In addition,the dynamic routing algorithm effectively reduces the impact of information loss caused by the pooling operation in the convolutional layer.The improved capsule network model has conducted coarse-grained sentiment classification experiments on data sets such as MR,IMDB,and CIN,and the accuracy rates are 80.12%,89.14%,and 76.57%,respectively.(2)For aspect-level sentiment analysis tasks,this thesis proposes a model based on BERT and joint attention mechanism.The BERT model is based on a multi-layer Transformer structure and semi-supervised pre-training through a large amount of corpus,which can provide a more robust and accurately described text representation.In this thesis,combining the characteristics of aspect-level sentiment analysis tasks,a downstream joint attention network is designed for the network structure of BERT.The core of the attention network is to achieve full attention interaction between contextual sentences,aspect words,and the pooled output of BERT to better extract key features in contextual sentences,thereby improving the effect of aspect-level sentiment analysis.The accuracy of the model on datasets such as Restaurant,Laptop and Twitter are 84.91%,78.36% and 75.43%,respectively.In addition,this thesis also explores the impact of the two BERT fine-tuning optimization methods on model performance,and verifies that the joint domain training method is effective.