Research On Emotion Analysis On Microblogging Text

With the popularity of the computer and Internet technology and the rapid development of Web2.0, social media network such as microblogging becomes the portal of information distribution, user emotion feedback and user commnunication. The analysis and tracking of emotions in microblogging text is helpful to understand users’ attention and emotional changes to specific product, person or event. Especially, it is helpful to provide essential information for real-time decision making. Thus, the researh on emotion analysis technique has great social and commercial value.This study investigates the techniques for identifying and classifying the emotions in microblogging text. The major works in this study including: 1. Considering the lack of emotion annotated corpus, in the study, a set of emotion annotation specification for microblogging text is designed. Following this specification, the microblogging level and sentence level emotion information are annotated. Up to now, 14,000 microbloggings which consists of 45,431 sentences are annotated. 2. Considering the fact that majority techniques based on inner-sentence features have the difficulty to achieve good performance on the short microblogging text with felxible expressions, a multi-lable microblogging emotion classification approach incorporating inner-sentence, intra-sentence and microblogging features is investigated. Here, the multi-label k-nearest neighbour classifier with inner-sentence features is devleoped as base classifier to generate the coarse classification results. The emotion transmission probabilities between neighbouring sentences and between the sentence and microblogging are further incorporated to update the sentence emotion classification results. Evaluations on NLP&CC2013 microblogging emotion classification datasets show that the proposed classification approach achieves 22.97% improvement on sentence level emotion classification performance from the baseline. 3. It is observed that the large number of microbloggings without emotions affects the emotion classification in real applications. Target to this problem, a two-stage emotion classification based on ensemble learning of multiple classifiers are inverstigated. Firstly, the classifiers based on GBDT and support vector machine are incorporated to determine whehter the microblogging has emotion expression. Secondly, multiple classifiers are incorporated in ensemble learning framework to classify the detected microbloggings with emotion expresssions. 4. Based on the above inverstigation, a system which monitoring the emotions in microbloggings for hot topics is developed to provide the analysis and visualization of emotions in microbloggings.The main contributions in this study are summarized as follows. Firstly, a emotion annotated corpus on Chinese microbloggings is constructed which is the largest avaliable resources in the world, based on our knowledge. Using this corpus as the golden standard, a large public evaluation on emotion classification technique is organized. Secondly, a coarse-fine microblogging emotion classification approach using intra-sentence feature, context feature and document feature is proposed. This approach achieves the hightest known performance on NLP&CC2013 microblog emotion analysis dataset. Thirdly, with the consideration on the characteristics of large-scale real microblogging text, a classification approach based on ensemble of multiple classifiers is designed. This approach improves the emotion classification on large-scale real text.