# 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/2306.08302v1/

## Key points

- Large language models (LLMs) like ChatGPT and GPT4 are impressive at generating text
- LLMs lack factual knowledge and interpretability
- Knowledge Graphs (KGs) store structured factual knowledge but struggle with generating new facts and representing unseen knowledge
- Unifying LLMs and KGs is of growing interest to leverage their respective advantages
- KG-enhanced LLMs: Incorporating KGs during pre-training and inference phases of LLMs to enhance understanding and improve inference capabilities.
- LLM-augmented KGs: Using LLMs to assist with various KG tasks such as embedding, completion, construction, graph-to-text generation, and question answering.
- Synergized LLMs + KGs: Emphasizing equal roles for both LLMs and KGs to enhance bidirectional reasoning driven by data and knowledge.
- Existing efforts within these three frameworks are reviewed in the article
- Future research directions aim to effectively combine the strengths of LLMs and KGs while addressing their individual limitations
- The goal is to create more powerful models that generate accurate and interpretable text while leveraging rich factual knowledge stored in KGs.

## 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 text. However, LLMs are often criticized for their lack of factual knowledge and interpretability. On the other hand, Knowledge Graphs (KGs) such as Wikipedia and Huapu store structured factual knowledge but face challenges in generating new facts and representing unseen knowledge. To address these limitations, there is a growing interest in unifying LLMs and KGs to leverage their respective advantages. This article presents a forward-looking roadmap for this unification, outlining 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. By integrating external knowledge from KGs, LLMs can enhance their understanding and improve inference capabilities. The second framework uses LLMs to assist with various KG tasks such as embedding, completion, construction, graph-to-text generation, and question answering. By leveraging the power of LLMs, KGs can benefit from improved performance in these tasks. The third framework emphasizes the equal roles played by both LLMs and KGs working together in a mutually beneficial way to enhance each other's capabilities for bidirectional reasoning driven by both data and knowledge. This article reviews existing efforts within these three frameworks and identifies future research directions with the goal of developing methods that effectively combine the strengths of LLMs and KGs while addressing their individual limitations. By unifying these two approaches, researchers aim to create more powerful models that can generate accurate and interpretable text while leveraging rich factual knowledge stored in KGs.