.
Abstract:
As a new form of visual language, emojis are widely used in social media for their vivid image and rich meaning.
Predicting the most likely emojis that fit a particular short text has become an important and challenging task in both academia and industry.
In this paper, we propose a hybrid sequence generation model, Seq2Emoji, to predict multiple emojis based on a short text.
Seq2Emoji is an encoder–decoder model, in which we consider the correlations between emojis and take the emoji prediction task as a sequence generation problem.
It extracts features through a hierarchical structure and self-attention mechanism and decodes them with a composite recurrent neural network before predicting emojis.
During the prediction, Diverse Beam Search algorithm is also introduced to increase the diversity of predicted emojis.
Experiments are carried out on our collected Weibo dataset (Chinese) and the results show that our proposed Seq2Emoji model is superior to the competitive models in both accuracy and diversity of emoji prediction.
.
https://www.sciencedirect.com/science/article/abs/pii/S095070512030856X
.
.
https://doi.org/10.1016/j.knosys.2020.106727
.
No comments:
Post a Comment