Attend to you: Real-time personalized headline generation for visual posts in social media

Recommending headlines tailored to users’ unique preferences significantly elevates their social media posting experience. This not only alleviates the frustration of struggling to craft compelling headlines but also helps organize their posting logics. However, previous works normally focus on summarizing and polishing static text posts that users have already finished, trying to come up with attractive or personalized headline recommendations. Put differently, they cannot generate personalized titles in real time, especially when they only have images from tweets to work with. Additionally, existing research lacks exploration into the role of users’ multimodal personalized information in headline generation. To address these gaps, we construct a comprehensive multimodal personalized dataset from RedNote for headline prediction and recommendation with users’ post history. Based on this dataset, we introduce a new task: Real-time Personalized Headlines Generation (RPHG) for visual posts, which focuses on generating preference-driven real-time headlines for RedNote posts before users draft their content. We also design two innovative metrics, RepQ-BLEU and FME-Score, to better assess personalization quality. Finally, we propose PHR, a multi-stage collaborative framework, which serves as a strong baseline for RPHG task. Extensive experiments demonstrate that the proposed metrics can effectively capture personalized headline quality within the RedNote context. Meanwhile, our proposed PHR significantly outperforms other baselines in the RPHG task for the platform.