Scholarlink

<p>Hashtags in microblogs serve as essential tools for content classification and recommendation, significantly enhancing the user experience. In modern microblogs, hashtags are often related to multimodal content, which typically includes both images and text. Besides, users’ historical posts provide valuable insights into their preferences and the contextual relevance of their content. Therefore, an effective hashtag recommendation algorithm must consider both the microblogs’ multimodal content and the information provided by users’ historical posts. Large Language Models (LLMs) have demonstrated remarkable capabilities in semantic understanding and reasoning, making them effective for various downstream tasks. To leverage the potential of modern LLMs, we propose <b>LLM-MHR2</b>, a novel approach that integrates historical information and multimodal microblog content to enhance multimodal hashtag recommendation performance. Specifically, the proposed Visual Prompt Former (VPF) module in LLM-MHR2 can effectively extract visual features and capture relationships from both current and historical visual inputs. In results, we evaluated LLM-MHR2 on the MACON dataset, where it outperformed the LLM-MHR while requiring only 11% of its trainable parameters. The results further establish its state-of-the-art (SOTA) performance, highlighting its efficiency in multimodal hashtag recommendation tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-MHR2: Leveraging Large Language Models for Multimodal and History-Aware Hashtag Recommendation

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Yuzhi Li,
  • Zhijie Tan,
  • Tuoyu Feng,
  • Tong Mo,
  • Weiping Li

摘要

Hashtags in microblogs serve as essential tools for content classification and recommendation, significantly enhancing the user experience. In modern microblogs, hashtags are often related to multimodal content, which typically includes both images and text. Besides, users’ historical posts provide valuable insights into their preferences and the contextual relevance of their content. Therefore, an effective hashtag recommendation algorithm must consider both the microblogs’ multimodal content and the information provided by users’ historical posts. Large Language Models (LLMs) have demonstrated remarkable capabilities in semantic understanding and reasoning, making them effective for various downstream tasks. To leverage the potential of modern LLMs, we propose LLM-MHR2, a novel approach that integrates historical information and multimodal microblog content to enhance multimodal hashtag recommendation performance. Specifically, the proposed Visual Prompt Former (VPF) module in LLM-MHR2 can effectively extract visual features and capture relationships from both current and historical visual inputs. In results, we evaluated LLM-MHR2 on the MACON dataset, where it outperformed the LLM-MHR while requiring only 11% of its trainable parameters. The results further establish its state-of-the-art (SOTA) performance, highlighting its efficiency in multimodal hashtag recommendation tasks.