# Unifying Large Language Models and Knowledge Graphs: A Roadmap

**Authors:** Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, Xindong Wu  
**arXiv:** [2306.08302](https://arxiv.org/abs/2306.08302)  
**Date:** 2023-06-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.08302v2/

## Key points

- Large language models (LLMs) like ChatGPT and GPT4 have gained attention in natural language processing and AI for their ability to generate human-like text.
- LLMs lack access to factual knowledge, as they are black-box models.
- Knowledge Graphs (KGs) provide structured representations of rich factual knowledge.
- KGs can enhance LLMs by offering external knowledge for inference and interpretability.
- Integrating LLMs and KGs can leverage their respective advantages.
- KG-enhanced LLMs: Incorporate KGs during pre-training and inference phases to enhance understanding of knowledge learned by LLMs through integration with external structured knowledge.
- LLM-augmented KGs: Utilize LLMs for various KG tasks (embedding, completion, construction, graph-to-text generation, question answering) to improve performance on those tasks by leveraging the capabilities of LLMs.
- Synergized LLMs + KGs: Emphasize equal participation from both sides to enhance bidirectional reasoning driven by data and knowledge, improving model performance and representation of knowledge in KGs.
- Future research directions include improving pre-training methods for incorporating external knowledge into LLM models, developing effective strategies for utilizing LLMs in KG-related tasks, investigating novel techniques for synergizing LLMs and KGs.
- Integration of LLMs and KGs holds potential for advancing natural language processing and AI by combining emergent ability & generalizability of former with rich factual information stored in latter.

## Summary

Large language models (LLMs) like ChatGPT and GPT4 have gained significant attention in the field of natural language processing and artificial intelligence due to their impressive ability to generate human-like text. However, one limitation of LLMs is their lack of access to factual knowledge, as they are essentially black-box models. On the other hand, Knowledge Graphs (KGs), such as Wikipedia and Huapu, provide structured representations of rich factual knowledge. KGs can enhance LLMs by offering external knowledge for inference and interpretability. The construction and evolution of KGs present challenges in generating new facts and representing unseen knowledge. Therefore, it is beneficial to integrate LLMs and KGs together to leverage their respective advantages. This article presents a forward-looking roadmap for the unification of LLMs and KGs. The proposed roadmap consists of three general frameworks: 1) KG-enhanced LLMs; 2) LLM-augmented KGs; 3) Synergized LLMs + KGs. The first framework involves incorporating KGs during the pre-training and inference phases of LLMs in order to enhance the understanding of knowledge learned by them through integration with external structured knowledge. The second framework utilizes LLMs for various KG tasks such as embedding, completion, construction, graph-to-text generation, and question answering in order to improve performance on those tasks by leveraging the capabilities of LLMs. Lastly, the third framework emphasizes equal participation from both sides in a mutually beneficial manner to enhance bidirectional reasoning driven by both data and knowledge which aims at improving both model performance as well as representation of knowledge in KGs. The article reviews existing efforts within these three frameworks and identifies future research directions including improving pre-training methods for incorporating external knowledge into LLM models; developing effective strategies for utilizing LLMs in KG related tasks; investigating novel techniques for synergizing LLMs and KGs etc.. In summary, the integration of LLMs and KGs holds great potential for advancing natural language processing and artificial intelligence by combining emergent ability & generalizability of former with rich factual information stored in latter thus enhancing performance & understanding complex language tasks better than ever before.