Multi-emotion Analysis Based On Deep Learning And Emoticon Distribution

In the background of big data,the emergence of various social media has brought about exponential growth of text data in the Internet.With the large amount of available data,more and more researchers have carried out related research work based on this large amount of data,such as the discovery research of key nodes in social networks,community discovery,textbased recommendation algorithm or sentiment analysis.However,the study of this article will analysis the emotions contained in short texts in a meticulously way,it is called text emotion analysis.In this article,according to the emotion classification research of American scholar Ekman,the emotions contained in the text can be classified into six categories: happy,sadness,angry,surprise,disgust,and fear.From the related research literature about emotion analysis,we found that the existing research work is mainly classify the sentiment of short text into two(positive,negative)or three categories(positive,negative and neutral).On the other hand,the related researches are seldom focus on emotion analysis.Furthermore,the previous work put the words and emoticons together or filter out the emoticons directly as noise when classifying the emotion of short text.Another important aspect is deep learning methods have shown good results in the text processing applications and there are not many deep learning methods cited into the text emotion analysis,hence,a new model has been proposed in this article to solve the emotion analysis task.The research work of this article are as follows:(1)Considering the importance of emoticons in the short text for emotion analysis.According to the distribution of emoticons in the data set and statistical analysis,the tendency of each emoticon to each emotion type were calculated.(2)Classification preprocessing of input text.Before embedding the text,the sentiment dictionary and emoticon dictionary are used to determine whether the input text contains explicit emotional words or emoticons,so that the input text is divided into four different sentence categories.Then,in the subsequent tasks,different fusion weights are learned according to different text categories to perform the fusion of emotion features to achieve a better emotion classification effect.(3)The emotion analysis model is proposed.Based on the research in(1)and(2)and combined with the deep learning method Bi-GRU network,this paper proposed the Semantic Emoticon Emotion Recognizer(SEER)model based on semantics and emoticons.The model use the Bi-GRU network and also apply the self-attention mechanism to extract the emotional features from the words aspect of input text more accurately,then,merge them with the emotional features of emoticons to complete the task of text emotion classification.In the end,the comparative experiments on the two data sets confirms the importance of emoticons in text sentiment analysis tasks,and also confirms the effectiveness of the new model proposed in this paper for text sentiment classification.