| With the development of network platforms,such as WeChat,Weibo,e-commerce,,individuals,companies and governments are increasingly inclined to publish important text information on the Internet.It is of great significance for both business and scholars to decipher the information containing in these texts.Fine-grained Sentiment,as an implicit piece of textual content,reflects the user's own emotions or their preferences for a particular topic.Accordingly,hot topics and social situations can be demonstrated clearly.Fine-grained Sentiment Analysis has great value in establishing evaluation system,obtaining real-time hotspots,controlling network public opinion and many other network information researches.Fine-grained sentiment analysis is a multi-task subject that combines multi-label topic prediction with multi-category emotion prediction.It needs to establish the theme of the research text first,and then to establish multiple emotional attributes for the theme.The problems of deep learning technology in fine-grained analysis of texts are the lack of sample numbers,complex multi-task calculations and insufficient semantic information.In the view of classical neural network model,this paper has three aspects for amelioration: multi-weight input,combining network and optimization,and realizes the application of deep learning technology in fine-grained Sentiment analysis.The main works are as follows:Aiming at the car comment dataset,a multi-weight model based on Softmax regression is proposed.In the model,the influence of text features and combination methods on topic classification is analyzed from the particularity of text and data.The enhanced text information is obtained by extending the text content,modifying the TF-IDF formula and adding the variance weight,and finally using the constructed shallow Layer neural networks to identify the subject of text.The experimental results show that the accuracy of model classification is 1.8% better than the traditional weighted Word2 vec model,and the accuracy of the topic prediction F1 is 64.5%.We analyze the problems of several classical neural networks in fine-grained classification,and propose a scheme of combining neural networks,which is an extension of the models used in the topic classification.(1)In order to avoid the influence of the sample of the subject classification misclassification on the sentiment classification,a label representation of the subject-emotional combination is proposed.(2)For the problem of semantic loss of multi-weight method in the network input layer,the input weighting method of the first layer network is improved.(3)In the course of the experiment,the problem of insufficient data samples was solved by adjusting the data enhancement scheme.The experimental results show that the model effectively performs fine-grained sentiment analysis on the car comment text. |
