Sentiment Analysis Of Microblogs With Rich Emoticons

With the rise and popularity of social media,people communicate and share information more and more frequently on social platforms.A large amount of information contains rich emotional tendencies,and emoticons are often used to express personal emotions,feelings and attitudes on social platforms.In recent years,with the development of mobile Internet,Sina Weibo has become a new social media platform facing the public and a social tool used by people every day.Due to the flexibility,randomness or diversity of Chinese microblog data,it is impossible to collect and sort out emotional information only by manual,and the posts on microblog contain a variety of content,such as emoticons and pictures,so how to quickly and accurately conduct emotion analysis has become a hot research direction in the field of natural language processing.Great progress has been made in the study of emojis and emotions.However,the current approach still has two major shortcomings: First,many researchers treat emoji as a single indicator,relying solely on artificial features or dictionaries for sentiment analysis.Second,they consider the emotion of emoticons and text separately,pay attention to their independence,and do not analyze the correlation between emoticons and text emotion enough.In order to better cope with these challenges,we established a Weibo sentiment analysis model based on Wnh Bert-Bi-LSTM,which can effectively evaluate the extent of emoticons’ effect on the emotional polarity of text,rather than just the emoticons themselves.By training phrase and emoji embedment in 280,000 Chinese microblogs,the model uses self-attention mechanism to evaluate the effect of emoticons on the polarity of text emotion,so as to better understand the changing trend of text emotion.By transforming emoticons into processable features,text messages and emoticons can be combined for analysis to explore the feature interaction between text messages and emoticons.To fully assess the impact of emojis on emotional polarity.Our model’s accuracy was 3.19% higher than the baseline models,as evaluated on 8,965 Sina microblog posts.Moreover,we created and released a new emoticon label corpus with more commonly used words and more comprehensive emoticon data than the existing corpus.