Sentence-level Emotion Classification

With the rapid development of the Internet Era, social networks such as Blog and Micro Blog have become a signification part of our life and entertainment. People are getting used to publish article, life information as well as their feelings on these social network platforms, which makes a great deal of text that is rich of emotion available on the Internet. The purpose of our research is to analyze the emotion expressed in these texts using machine learning method, such as, happy, anger and sorrow. In recent years research on emotion classification has gain more and more attention from scholars in the field of computer linguistics, making itself a fundamental task in the natural language processing community. Emotion classification research is categorized into several categories including document-level, sentence-level(short text) and word-level. The topic of this paper is targeted at the sentence-level emotion classification.First, we propose a joint learning method of emotion and sentiment based on integer linear programming. We use the labeled data of both emotion classification task and sentiment classification task to explore the classification of the join of two kinds of tasks. Specifically, we designed a series of restrictions between sentiment orientation and personal emotion based on the connection between sentiment orientation and personal emotions. Then we use integer linear programming to produce a refined result of emotion and mood classification. As demonstrated in the experiment, this method is beneficial for both emotion classification and sentiment classification tasks.Second, we propose a sentence-level emotion classification method based on label dependence. This method produces the relationship between sentiment labels by studying the co-occurrence phenomenon of sentiment label and the number of single sample sentiment labels and then use this relationship to process a multiple-label sentence-level sentiment classifying. Specifically, we use the label dependence to set up the restriction between sentiment labels and use integer linear programming to produce a refined result of multiple-label sentence-level emotion classification. Experiment shows that this method can effectively improve the performances by using the relationship between labels to obtain a better classification result.Third, we propose a context-based sentiment classification method. The idea of this method is to employ the context dependence to multi-label sentence level sentiment. Specifically, we construct a false sample for each sentiment label along with the sample network of these constructed false samples against every sentence instance to describe the context dependence. Based on that, we construct a relational factor graph model to perform context relationship sentiment label learning based on the original context features. Empirical studies demonstrate that the proposed method effectively improves multi-label sentence-level sentiment classification by using the sample context dependence.