Research And Analysis In Fine-grained Sentiment Of Text Based On Deep Learning

The increasing number of Internet users has caused a massive growth of text data,as well as the diversity and complexity of the words used in the commentary text.At this time,it is obviously not feasible to rely solely on expert experience,or manual retrieval to deal with these massive amounts of data,so it is particularly important to introduce more intelligent natural language processing technology to solve this problem.With further research,sentiment analysis is widely used in text processing.Based on deep learning and joint learning,this thesis makes the following studies in view of the problems existing in fine-grained emotion analysis.(1)CNN can only obtain the dependence between local words,not the overall structure information of the sentence,and ignore the modeling of the relationship between the target word and the text context,which makes the existing methods still have some problems.To solve these problems,this thesis introduces a multihead attention mechanism into the GCAE model,while adding the modeling of specific aspect sequences and learning the interaction relationship between them during training,and then proposes the GCAE-MHA model(GCAE based on Multi-Head Attention mechanism model).The model can simultaneously learn the relationship between the context and the specific aspect during training,and is used to capture the information in the context sequence that is closely related to the specific aspect,thereby increasing the modeling of the specific aspect and making up for the original GCAE model only modeling the context information.(2)Processing the extraction of emotional target objects and the discrimination of their emotional polarity separately has led to training and learning methods that focus only on an individual model and ignores some hidden information that may improve the target task in the relevant task while processing.Therefore,this thesis introduces a more advanced pre-training BERT model and then proposes the GL-IJM model(Global and Local Information and Joint Model),which is composed of sharing layer,information fusion layer and output layer.Based on multi-tasking learning,the model effectively utilizes the powerful feature extraction ability of the BERT pre-training model,while jointly processes ATE with SPC using a multi-task learning framework.In addition,the model integrates local information and global information through the carefully designed information fusion model IFA,as a feature of sentiment classification,which effectively improves the effect of total task of sentiment analysis.(3)Based on sentiment analysis and data visualization,this thesis investigated the emotional situation of public opinion during the COVID-19 period in China,and timely analyzed the main reasons for the changes in people’s emotional polarity during the epidemic period.In addition,combined with the innovation of sentiment analysis in this thesis,the emotional classification system module is designed for the public opinion during the COVID-19 period.This module can effectively classify the public opinion data of netizen during the epidemic period,and accurately make the emotional discrimination of the specific entities of the network public opinion text.Therefore,users can more timely and effectively grasp the status of public opinion ecosystem and enhance the management of false public opinion and grasp the law of public opinion evolution from multiple perspectives.