Research On Emotion Analysis On Chinese Microblog

With the rapid development of the Internet and the popularity of mobile intelligent terminal, more and more users express their opinions and feelings through the social networks, generating a large number of rich emotional expression text. Recently, sentiment analysis has attracted many researchers' attention in the research community of computational linguistics and becomes a basic hot topic. Emotion analysis is a basic task of sentiment analysis, which aims to automatic analyze involving emotion(e.g. happiness, angry, sadness, fear) towards a piece of text. This paper attempts to study emotion analysis on Chinese microblog, and researches the following aspects of emotion analysis:First, this paper proposes a novel emotion taxonomy which contains both basic emotions and complex emotions for annotating emotions expressed in microblog text, and builds a large Chinese microblog emotion corpus. Meanwhile, in order to reduce the workload of annotation, this paper puts forward a method that combines automatic annotation, which improves the efficiency of annotation. The final annotation results show that our proposed emotion taxonomy can be better distinguish the different emotion categories, and achieve a much higher annotation consistency.Second, this paper proposes a novel emotion recognition approach based on syntactic information. The method is supervised and regards emotion recognition as a binary classification problem. We leverage POS(part of speech) sequence and syntactic tree to generate POS sequence patterns, rewrite rules and bigrams of syntactic labels as features for text representation, and then employ them in a machine learning classification algorithm. This method effectively overcomes the disadvantages of microblog text short and the relatively small information. Experimental studies show that our approach is very effective for emotion recognition.Third, as for emotion category imbalanced distribution, we propose an emotion classification with ensemble learning approach. The main idea is to use the under-sampling approach to generate a set of training subsets. These subsets are then employed to train multiple sub-classifiers. We propose an ensemble learning method to combine these sub-classifiers. Experimental studies demonstrate that our proposed approach can take full advantage of more training sample to improve performances of imbalanced emotion classification.