# NoteLLM: A Retrievable Large Language Model for Note Recommendation

**Authors:** Chao Zhang, Shiwei Wu, Haoxin Zhang, Tong Xu, Yan Gao, Yao Hu, Enhong Chen  
**arXiv:** [2403.01744](https://arxiv.org/abs/2403.01744)  
**Date:** 2024-03-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.01744v1/

## Key points

- Authors emphasize the importance of recommending notes aligned with user interests in online communities
- Limitations of existing methods that rely solely on BERT-based models for note embeddings
- Introduction of a novel unified framework called NoteLLM for item-to-item (I2I) note recommendation
- Utilization of Note Compression Prompt to condense notes into a single special token and contrastive learning approach for related note embeddings
- Automatic generation of hashtags/categories through instruction tuning
- Validation demonstrating the effectiveness of the proposed method compared to online baselines, with significant improvements in Xiaohongshu's recommendation system
- Enhancement of note embeddings by incorporating important cues and leveraging LLMs' superior performance in understanding natural languages

## Summary

In their paper titled "NoteLLM: A Retrievable Large Language Model for Note Recommendation," authors Chao Zhang, Shiwei Wu, Haoxin Zhang, Tong Xu, Yan Gao, Yao Hu, and Enhong Chen address the importance of recommending notes aligned with user interests within online communities. They highlight the limitations of existing methods that solely rely on BERT-based models to generate note embeddings for assessing similarity. These methods may overlook crucial cues such as hashtags or categories, which encapsulate the key concepts of notes. By learning to generate hashtags/categories and leveraging Large Language Models (LLMs) over BERT, the authors introduce a novel unified framework called NoteLLM for item-to-item (I2I) note recommendation. The  framework utilizes a Note Compression Prompt to condense a note into a single special token and employs a contrastive learning approach to learn embeddings of potentially related notes. Additionally,  is used to summarize notes and automatically generate hashtags/categories through instruction tuning. Extensive validations conducted on real scenarios demonstrate the effectiveness of the proposed method compared to online baselines, showcasing significant improvements in Xiaohongshu's recommendation system. This innovative approach not only enhances note embeddings by incorporating important cues but also leverages LLMs' superior performance in understanding natural languages. The authors' research sheds light on the potential of utilizing advanced language models for personalized note recommendations tailored to individual user preferences within online communities.