面向社交媒体的分级注意力表情符预测模型

Hierarchical and attentional emoji prediction model for social media

将预测社交媒体表情符的任务作为文本分类问题,将输入文本映射到最有可能的伴随表情符号。首先,通过研究帖子中出现的表情符与标签之间的关系,提出一个基于标签、发帖用户、发帖时间、发帖地点的注意力机制;其次,添加表情符位置特征;最后,探讨注意力机制、分级模型对于表情符预测任务的作用,训练多种模型并比较其预测效果。实验结果表明,模型对于不同使用频率的表情符的预测效果均有显著提升,模型是可行的、高效的。

This paper treated the task of social media emoji prediction as a text classification problem,and mapped the input text to the most likely accompanying emojis.Firstly,it proposed an attention mechanism based on hashtags,posting users,posting time,and posting location by studying the relation between emojis and hashtags appearing in the posts.Secondly,it added the emoji position feature.Finally,it discussed the attention mechanism and hierarchical model for the role of emoji prediction task,and trained the various models to compare their prediction effects.The experimental results show that the model has significant improvement on the prediction effect of emojis with different frequency of use.The model is feasible and efficient.