Genre-based Emoji Usage Analysis and Prediction in Video Comments

摘要

Emojis can be seen as a visual language inserted in texts to express emotions, attitudes, and situations. It is widely used in computer-mediated communication, e.g., video comments. The emojis can express more detailed information beyond text information, and their usage can improve interlocutors’ communication efficiency and emotions. The latest advances in natural language processing and deep learning have made it possible for chatbots to automatically add emojis in their dialogue. Precisely predicting emojis to be added is very challenging, especially in the video comments, where the use of emojis is complex, subtle, and associated with the cultural characteristics of video genres, e.g., anime and dancing. In this article, we first construct a benchmark dataset Bilibili comments dataset with more than 3.9 million comments that contain emojis in the video-sharing website Bilibili and then statistically analyze features of emoji’s usage of video genre, comment content, and where an emoji appears in a sentence. According to the analyzed results and the gated recurrent unit (GRU) neural network, we propose a novel model of genre-based multitask GRU (GM-GRU) and its attention-added edition (GM- GRU+) to predict an emoji’s category and position in a video comment. Our experiment and evaluation show that the proposed method can significantly increase the accuracy of predicted emojis for video comments.