Emotion Wheel And Lexicon Based Text Emotion Distribution Label Enhancement Method

With the rapid development of new social media and mobile Internet,more and more people express their emotions or opinions by making comments on social networks.As an important research topic in the field of natural language processing,sentiment analysis technology can effectively analyze a large amount of text information generated on the Internet and discriminate the views and emotional tendencies of netizens,which has attracted more and more attention from researchers.Traditional text emotion analysis mostly assumes that each sentence has only one or more associated emotion labels,and can recognize which kinds of emotions are contained in the target sentence,but it is not able to quantitatively answer the specific degree of expression of each related emotion.This problem can be solved by Emotion Distribution Learning(EDL)to predict the specific proportions of all the emotions contained in a sentence.EDL is a recently proposed effective multi-emotion analysis model,whose key idea is to handle the emotion fuzziness by associating each instance with an emotion distribution.Each component in an emotion distribution is the expression degree of the corresponding emotion on the given instance.Different from the traditional single label learning or multi label learning,EDL can quantitatively model multiple emotions simultaneously,which has obvious advantages in solving tasks with emotion fuzziness.Nowadays,one of the most critical difficulties of EDL is the lack of emotion distribution marked text datasets.Utilizing the existed single-label emotion datasets in EDL is a possible way to solve this problem,where emotion distribution label enhancement methods can be applied to convert the instances' emotion label to emotion distribution.This paper proposes an Emotion Wheel and Lexicon based emotion distribution Label Enhancement(EWLLE)method by extracting the affective words' emotional information and introducing the prior knowledge of the Plutchik's emotion wheel psychological model.Based on the psychological emotion distances,the EWLLE method generates the discrete Gaussian distributions for the emotion label of the sentence and the emotion labels of affective words respectively.Then,the two kinds of distribution are superposed into a unified emotion distribution.Different from the existing emotion distribution enhancement methods,the EWLLE method takes into account both the psychological knowledge of emotion and the linguistic information of affective words in addition to traditional supervised information(emotion labels of sentences).Extensive comparative experiments on 7 commonly used Chinese and English text emotion datasets show that the proposed EWLLE method is superior to the existed emotion distribution label enhancement methods in the emotion recognition task.