Emoji Embedded Representation Based On Emotion Distribution

With the development of social networks and intelligent terminals,social media has been widely used as a medium for disseminating information.Sentiment analysis is the process of collecting and analyzing people’s opinions,thoughts and impressions on various topics,products,themes and services.Sentiment analysis of text data such as user comments on social media can enable people to communicate,interact,and make decisions effectively.Emoji is a non-text language that contains rich emotional meanings and is based on image symbols,filling the need for nonverbal information in social media.It can help users express emotions and improve communication efficiency.It is often used in conjunction with text in social media to express users’ emotional intentions.Considering emoji information in social media oriented sentiment analysis models can effectively improve the generalization performance of the model.Many natural language processing applications currently used for social media rely on presentation learning.Constructing effective emoji representations is the basis for sentiment analysis models to consider emoji information.At present,most of the existing emoji representation methods use traditional word embedding models,mainly consisting of two types: emoji embedded representation based on description text and emoji embedded representation based on model training.Both representations extract emoji embedded vectors containing semantic information and low-dimensional density based on the co-occurrence pattern in the training data.However,traditional word embedding models do not consider emotional information and cannot directly express the degree of association between emoji and fine-grained emotions in embedded vectors.Based on the feature that emoji can simultaneously express multiple emotions with different intensities,and drawing on the research ideas of emotional distribution,this paper proposes an emoji embedded representation method based on emotion distribution.This paper uses the soft labels of BERT-based sentiment prediction model to learn emoji embedded representations directly associated with fine-grained emotions from real social media data.The effective emoji embedded representations are learned while the sentiment prediction model is trained,and the degree of emoji expression on various fine-grained sentiments is directly modeled by emotion distribution,so that the embedded representations contain multiple sentiment information of emoji.Three comparative experiments were conducted on a Chinese Weibo dataset containing emoji:vector representation of emoji for visualization experiments,emoji representation sentiment polarity consistency detection experiment and emoji representation mapping fine-grained emotion comparison experiment.The experiments show that the proposed method can effectively learn emoji embedded representations directly related to finegrained emotions.The learned emoji embedded representations can describe the emotion relations among emoji with different polarities while containing accurate emotion information,and build an emoji representation space with high quality of emotion expressions,so that the emoji in the representation space have better emotion consistency between emoji close to each other.