Dimensional Based Emotion Recognition Using Expand Word And Modifier Structure

Affective computing is a hot research direction in the field of natural language processing.With the increase of the amount of the network information, the emotional computing with huge amount of text has become an important branch of research in big data field. Most of previous studies is classification of emotion, they divided emotion into different categories, then predict text’s emotion belong to which advanced category. However it is hard to decide how many emotion categories there are. So in this paper we use a multi-dimensional model to quantify emotion. In this model the text’s emotion is converted to two consecutive scores which are Valence(the degree of positive and negative) and Arousal(the degree of calm and excitement).Through this method emotion can be mapped onto a two-dimensional plane, thus we can avoid the problem that to define the category of emotion accurately.In this paper we use lexicon method to calculate the value of text’s Valence-Arousal. The lexicon contains a list of emotion words which have been marked with Valence-Arousal. Then we use the emotion lexicon to recognize the emotion word in text, after that we can predict the text’s Valence-Arousal using the Valence-Arousal of the emotion word in the text. However there are two major drawbacks: First, the scale of lexicon is restricted and the cost of large number of artificial marking is too high. Second, the emotional word appear together with modifiers such as “happy”, when it co-occurrence with “not” the word’s Valence and Arousal values vary greatly. To solve these problems we propose a regression model based on word similarity to amplify the emotion lexicon automatically, it solve the problem of lexicon with limited emotional words. And we build a quantified impact model for modifier to calculate the influence of modifier on the value of emotional words.In the absence of Chinese Valence-Arousal lexicon and corpus, we take a lot of time to complete a lexicon which contains 1653 emotional word with Valence-Arousal score, and a corpus which contains 720 texts as the basis for the study.