Joint Modeling Of Text Sentiment And Topic For Social Media

The development of network technology and the social of netizens needs have made varieties of social media spring up rapidly.More and more people are using various social networks such as Microblog,BBS,and so on,to participate in various topics,express their feelings and opinions,and learn from others' feelings and opinions.As the most popular social platform,Microblog is more informative,more broadly used,spreading faster,and more dynamic and real-time compared to traditional network media.Hundreds of millions of microblogs are constantly posted and spreading every day.On the one hand,the massive data have rich sentiment and topic information,which contains huge economic,social,and scientific value;on the other hand,the data is informal and feature sparse,which makes sentiment and topic analysis so difficult.This paper studies text sentiment and topic analysis for social media.It highlights optimizations of sentiment and topic analysis task with bias,sentiment intensity,and general knowledge one after another,considering social media texts as datasets.It includes three contributions:Firstly,this paper proposes the concept of bias including subjective bias and objective bias,and constructs Bias-Sentiment-Topic(BST)model based on it.BST considers that the sentiment and topic generation of text have different dependencies under the different conditions of biases.It adds a bias layer based on the Joint Sentiment-Topic(JST)model and Reverse Joint Sentiment-Topic(Reverse-JST)model and incorporates the prior knowledge of bias,sentiment,and topic,to achieve the joint modeling of the three features.BST takes advantage of the bidirectional dependency relationships between sentiment and topic generation.As a result,it can not only improve the performance of text sentiment and topic analysis,but further mine the bias knowledge besides the sentiment and topic knowledge.Secondly,this paper extends sentiment into sentiment category and sentiment intensity,and constructs Bias-Sentiment category-sentiment Intensity-Topic(BSIT)model.BSIT considers that the generation of sentiment intensity depends on sentiment category or topic under the different conditions of biases.It adds a sentiment intensity layer based on the BST model and incorporates the prior knowledge of bias,sentiment category,sentiment intensity,and topic,to achieve the joint modeling of the four features.BSIT takes full advantage of sentiment category,sentiment intensity,and topic under different biases as well as their relations to implement the longitudinally deep analysis of text sentiment.Thirdly,this paper constructs an improved Gibbs sampler based on the generalized Pólya urn(GPU)model.The sampler takes the word embedding set of words trained from massive text data as general knowledge,and incorporates the knowledge into the Gibbs sampling process of BST and BSIT based on the GPU model.This improved sampler use the general knowledge to perfect the semantics in test sets consisting of social media texts,which further enhance the performance of text sentiment and topic analysis for social media.In summary,this paper makes a further study of text sentiment and topic analysis for social media from the three aspects above.In addition,it carries out a series of comparative experiments based on the Twitter text datasets.The experimental results demonstrate that the proposed methods not only improve the performance of the text sentiment classification for social media,but also improve the performance of the topic classification effectively.