Article Preview
Top1. Introduction
With the rapid development of machine learning, more and more projects are gradually focusing on everyday life. Machine learning has achieved a lot in fraud detection (Almomani, A., et al., 2022; Gaurav, A., et al., 2022), traffic management (Liu, R., et al., 2022; Jain, A., et al., 2022), and even big data (Stergiou, C., et al., 2021; Lv, L., et al., 2022; Xu, Z., et al., 2023; Barbosa, A., et al., 2022). In the field of natural language processing, Yen et al. (2021) identify social user status by analyzing post content and social behavior. Gu et al. (2022) utilize a contextual Word2Vec model to understand the words out of vocabulary in Weibo corpus. Sentiment analysis of social users' speech can reveal users' views on current hot events and control public emotions. Alowibdi et al. (2021) discovered that COVID-19 has propagated fear, anxiety, hope, and other emotions throughout the general population by examining the content of Twitter users' messages. In order to identify conflicts in trending Twitter topics, Al-Ayyoub et al. (2018) suggest a hybrid strategy. This is done in order to examine the dynamics of online groups and the behavior of their interactions. The crucial factors that influence tourists' pleasure were identified by Bai et al. (2023) by examining the comments posted by tourists. This discovery is extremely important for the study of consumer behavior. More and more social media data is distributed in unstructured forms (Singh, S. K., & Sachan, M. K., 2021), such as emoji, which can help sentiment analysis. As a kind of emotional symbols, emojis are frequently used with natural language, especially in social media posts such as Weibo(in China) and Twitter. There are thousands of icons that can express emotions. And there are many variants of emojis that have similar meaning but have different features such as skin tones and image amounts, etc. Still, this does not affect how well emojis combined with short text can express the emotions of users. Evans (2017) sees nonverbal cues as an expression of emotions. However, in traditional digital communication, these cues may get lost, which can lead to communication bias. Nevertheless, emojis fulfill this function. To some extent, emojis even reflect the real emotional polarity of the authors. It may lead to the opposite emotional polarity via text without emojis. According to Na’aman et al. (2017), emojis can serve as syntactic components in the text in the same way words do. But on the other side, emojis may also mislead humans. Each individual has a different perception of emotions containing emojis, which is known as cognitive bias (Miller, H., et al., 2017). Luckily, short text combined with emojis can express emotions that fit what most people perceive. The applications of emoji prediction are far-reaching. Social media platforms can analyze user opinions on hot topics through the sentiment of social posts, while commercial platforms can improve product quality and service experience based on user feedback. In addition, for intelligent question-answering(QA) models like ChatGPT, the abstract understanding and actual use of emoji can also improve the performance of the model on more tasks. Thus, prediction of emojis would be beneficial for achieving better language understanding (Barbieri, F., Ballesteros, M., & Saggion, H., 2017).
The emoji prediction task was originally proposed in (Kralj Novak, P., 2015). The task aims to find the most appropriate emoji according to a short piece of text. As everyone knows, it has become a trend that more and more natural language processing(NLP) task datasets contain emojis. In general, it would be helpful to pre-train models by predicting emojis to boost their performance on other NLP tasks. And the knowledge learned via emoji prediction tasks can be well transferred to solve other NLP tasks e.g., emotion prediction, sentiment analysis and sarcasm detection (Felbo, B., 2017). The fine-tuned models pretrained on emoji prediction datasets tend to perform better.