IEEE VIS 2024 Content: Bavisitter: Integrating Design Guidelines into Large Language Models for Visualization Authoring

Bavisitter: Integrating Design Guidelines into Large Language Models for Visualization Authoring



Jiwon Choi - Sungkyunkwan University, Suwon, Korea, Republic of

 Jaeung Lee - Sungkyunkwan University, Suwon, Korea, Republic of

 Jaemin Jo - Sungkyunkwan University, Suwon, Korea, Republic of

 Download Supplemental Material

 Room: Bayshore VI

2024-10-17T18:39:00ZGMT-0600Change your timezone on the schedule page
2024-10-17T18:39:00Z

Exemplar figure, described by caption below — Bavisitter’s visualization authoring workflow. A) The user requests a visualization to an LLM by prompting “Show me the average yield by site.” B) The LLM generates an ineffective visualization design that uses a connection mark to encode the categorical attribute on the x-axis. C) Bavisitter detects the design issue in the generated visualization and gives feedback to the LLM by modifying the original prompt, e.g., appending “Change mark to bar”. As a result, the user can author visualization designs that conform to known design guidelines and knowledge while exploiting the flexibility that the LLM provides.

Fast forward

Keywords

Automated Visualization, Visualization Tools, Large Language Model.

Abstract

Large Language Models (LLMs) have demonstrated remarkable versatility in visualization authoring, but often generate suboptimal designs that are invalid or fail to adhere to design guidelines for effective visualization. We present Bavisitter, a natural language interface that integrates established visualization design guidelines into LLMs.Based on our survey on the design issues in LLM-generated visualizations, Bavisitter monitors the generated visualizations during a visualization authoring dialogue to detect an issue. When an issue is detected, it intervenes in the dialogue, suggesting possible solutions to the issue by modifying the prompts. We also demonstrate two use cases where Bavisitter detects and resolves design issues from the actual LLM-generated visualizations.