# A Preliminary Roadmap for LLMs as Assistants in Exploring, Analyzing, and Visualizing Knowledge Graphs

**Authors:** Harry Li, Gabriel Appleby, Ashley Suh  
**arXiv:** [2404.01425](https://arxiv.org/abs/2404.01425)  
**Date:** 2024-04-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.01425v1/

## Key points

- Li, Appleby, and Suh study the potential of large language models (LLMs) in assisting users with visual exploration and analysis of knowledge graphs (KGs).
- Mixed-methods approach involving surveys and interviews with 20 professionals from industry, government laboratories, and academia.
- Participants desire LLMs to aid in data retrieval from KGs through collaborative query construction.
- Interest in using LLMs for uncovering relationships within KGs through multi-turn conversations and generating on-demand visualizations.
- Preference for a chat-based 'widget' integrated into existing workflows to interact with LLMs through visualizations.
- Participants prefer a combination of annotated visuals alongside summarizing text when viewing LLM outputs.
- Concerns about maintaining semantic intent when translating natural language questions into KG queries.
- Researchers propose a preliminary roadmap for designing LLM-driven systems for exploring knowledge graphs, outlining future opportunities in this design space.

## Summary

In their study, Li, Appleby, and Suh explore the potential of large language models (LLMs) in assisting users with the visual exploration and analysis of knowledge graphs (KGs). Through a mixed-methods approach involving surveys and interviews with 20 professionals from various sectors, including industry, government laboratories, and academia, the researchers identified key preferences and concerns regarding the integration of LLMs in KG exploration. The findings reveal a strong desire among participants for LLMs to aid in data retrieval from KGs through collaborative query construction. Additionally, participants expressed interest in using LLMs to uncover interesting relationships within KGs through multi-turn conversations and to generate on-demand visualizations that enhance trust in the model's outputs. To facilitate interactions with LLMs, participants favored a chat-based 'widget' integrated into their existing analysis workflows, allowing them to guide the model through interactions with visualizations. When it comes to viewing LLM outputs, participants preferred a combination of annotated visuals (such as subgraphs or tables extracted from KGs) alongside summarizing text. However, concerns were raised about the model's ability to maintain semantic intent when translating natural language questions into KG queries. Based on their analysis of interviews, the researchers propose a preliminary roadmap for designing LLM-driven systems for exploring knowledge graphs. This roadmap outlines future opportunities in this emerging design space and provides valuable insights for further research and development in leveraging LLMs as assistants in KG exploration and visualization.