Sentiment Enhancement Method Based On Global Semantic Learning And Its Applications In Text Sentiment Analysis

with the widespread applications of Internet technology,a great number of user-generated contents constantly appear on the network.Through fully mining the valuable information contained in these contents by sentiment analysis can provide new vitality for every industry.As an important foundation work of text sentiment analysis,sentiment classification has attracted more and more attention.The accuracy of sentiment classification result is not only determined by the structure designing of classification model,but also affected by the data set.Apparently,the training corpus with unapparent or conflicting emotional features will introduce new convergence errors into the process of model training.Ultimately,the convergence error will affect the effect of text sentiment classification.Because it can weaken model's ability to capture and discern emotional features.For this reason,this dissertation focuses on the sentiment enhancement of corpus with weak sentiment tendency,and aims to improve the effect of text sentiment classification.The specific work mainly includes the following three aspects:(1)Dividing method of strong and weak sentiment tendency corpus.First,the LSTM sentiment intensity calculation model based on residual connection was constructed to improve the accuracy of predicting sentiment intensity and alleviate the degradation caused by deep networks.Then,combining some parameters from the model constructed in the previous step to design the calculation method about sentiment intensity based on the transfer strategy of parameters sharing.It is used to calculate intensity of text corpus in sentiment classification dataset.On this basis of above,designing the method of corpus dividing based on sentiment intensity to get strong or weak sentiment tendency corpus.(2)Text sentiment enhancement method based on global semantic learning.The encoding process of Variational Auto-Encoder was improved by designing and merging maximum extraction,mean extraction and document information vector.Then the Variational Auto-Encoder for Global Semantic Learning was constructed,which re-encoded the weak sentiment corpus by means of learning the language sequence features and emotional semantic features of the strong sentiment corpus.It can solve the issues of emotional features conflicting or being unapparent and achieve text sentiment enhancement.Eventually,verifying whether sentiment enhancement is effective in improving the effect of sentiment classification through experiments.(3)The realization and application of text sentiment analysis system.Based on the research achievements,a text sentiment analysis system was designed and developed in the form of B/S architecture,and the application in predicting the trend of stock price was completed.From the perspective of training data and training process of deep model,we proved that the quality of text corpus is also an important factor of affecting the classification accuracy in the study of text sentiment classification.Meanwhile,through designing a sentiment enhancement method based on global semantic learning,it provides an effective way to improve the quality of text corpus and the effect of sentiment classification.