In new social platforms, users usually use hashtags to mark the keywords or topics of the posts when posting posts, which will increase their participation in social media. In this article, considering that the text of the users post can better express the users own thoughts, a hierarchical structure is used to extract text features from the three levels of words, phrases, and sentences, and propose a summary attention mechanism for the text content. The semantic content of each level is summarized as a feature vector, and then a text-enhanced common attention model is proposed to merge the semantics of each level with the image modal. At the same time, considering that different users have different hashtag preferences, an external storage unit is introduce to record the historical hashtag habits of each user, calculate the similarity influence vector between the current post to be recommended and the historical post, and establish the user personalized module. The overall hashtag recommendation results are generated based on the analysis of multi-modal post content understanding and personalized modules. Experimental results on real data sets show that our model has a great improvement in accuracy, recall and F1 score compared with other models, the two attention mechanisms for multi-modal content understanding and the users personalized modeling proposed in this paper all contribute significantly to the overall recommendation effect.

text hierarchical modeling; co-attention mechanism; text attention mechanism; multimodal recommendation; personalized recommendation