Sentiment Analysis On Short Text In Social Media

Sentiment Analysis(SA) focuses on finding sentiment information from huge texts and predicting the trend of sentiment in the text. Sentiment Analysis plays an important role in data mining tasks for the Internet.Microblog Sentiment Analysis(MSA) is a popular and important theme in social networks. Microblog platform such as Twitter can collect rich microblog messages everyday, and such messages can be used in all kinds of applications in the Internet. In this paper, we propose two new models in order to improve the performance on MSA.Firstly, we propose a novel semi-supervised learning approach for MSA. Specifically, we make use of microblog-microblog relations to build a graph-based semi-supervised classifier. We incorporate both social relations and text similarities into building microblog-microblog relations. Our model connects labeled data and unlabeled data via microblogmicroblog relations. Experiments on two real-world datasets show that our graph-based semi-supervised model outperforms the existing state-ofthe-art models.Secondly, MSA tasks often focus on some specific domain. It is typically diffcult or time-consuming to collect suffcient domain-specific sentiment labels for training the MSA models. Moreover, there are some domain-specific words whose sentiment values are diffcult to identify.To solve these problems, we propose a novel transfer learning approach for MSA in this paper. Our approach is motivated by the following fact: there are informative domain-independent sentiment dictionaries which can be easily obtained and we can transfer such domain-independent lexical knowledge into domain-specific sentiment analysis. More specifically,we exploit sentiment dictionaries and make use of word-word relations to build a graph-based transfer learning classifier for MSA. Experiments on two real-world datasets show that our transfer learning model can outperform existing state-of-the-art models in MSA.