Semi-Supervised Chinese Text Emotion Classification Based On Manifold Regularization And Emotional Factor

The Internet has become a large plat form for people to exchange viewpoints and express emotion, which results in a great amount of subjective text on the internet. In terms of investigating public attitudes for the government or customer opinions for the business, tracing hot issues and the like, analysis of subjective text is of great practical significance. This study focuses on the Chinese text emotion classification problem, trying to figure out appropriate mathematical models and useful knowledge for emotion classification.Over the past decade, Chinese text emotion classification has attracted interests of many researchers in the field, leading to considerable achievement, but also exposing many problems listed below:First, there is no enough labeled emotion text, which to a large extent restricts the research of text-based emotion classification. Time-consuming and low-consistency of emotion text annotation are the main reasons here, making it more difficult to get labeled corpus than that in segmentation or POS tagging tasks.Second, emotion classification of high-quality requires more deep language information and emotional knowledge for the superficial text features are no longer able to distinguish the emotion category effectively.Finally, subjectivity and individuality, as the main characteristics of emotion, introduce high emotional ambiguity of the text as well as great challenge to the automatic computation.Against the above problems, we proposed semi-supervised learning method for text-based emotion classification, and explored relative knowledge to apply to emotion classification. The main contribution of this thesis can be summarized as follows:1. A general framework of semi-supervised probability discriminant model based on manifold regularization is proposed. Applying the framework, we implement the manifold regularization-based semi-supervised Multivariate Logistic regression (i.e. MRMLR algorithm) for emotion classification. Theoretical analysis and experimental results prove the effectiveness and stability of MRMLR algorithm.2. We figure out a transductive learning algorithm (i.e. MRTRU algorithm) based on manifold regularization for emotion classification. The algorithm uses a variation of the expectation maximization algorithm (EM) to solve the optimization problem of the above framework, which does not need make any assumption about the specific form of the data distribution, and get out of the restriction that samples must be represented in the feature vector space. Experiments show that the proposed approach outperforms the state-of-the-art graph-based transductive learning methods.3. We do research on emotional knowledge and apply it to emotion classification. An emotional knowledge base is constructed, which provides distribution of emotion reflected by words over different types of emotional factors (including incentives, emotional experience, behavior and consequences, and the outward manifestation). We adopt rule-based methods to use it in emotion classification. Experiments results prove the rationality of the proposed emotional knowledge system, as well as the effectiveness of the knowledge base.4. We propose the semi-supervised emotion classification methods combined with emotional knowledge to overcome the problems caused by the limited number of knowledge and unlisted words. We feed the result obtained by using the rule-based method as the priori into statistical models. The experiments show that both the advantages of emotional knowledge and statistical methods are preserved, making the performance of emotion classification improved significantly.