InReAcTable: LLM-powered Interactive Visual Data Story Construction from Tabular Data
summary
A framework and prototype called InReAcTable (In-RACTIVE) combines an Acting module with insight graph structural filtering and a Reasoning module with retrieval-augmented LLM generation to recommend related insights and assemble tree-structured visual data stories from tables. The system extracts insights, narrows candidates via graph search, then retrieves top-K textual descriptions for LLM semantic ranking. Demo uses a 1,280-value game-console sales table and the prototype generates coherent story trees; a user study with 18 participants vs AWS Quicksight and VisGBT reports improved story construction and positive usability feedback.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 1:46Challenges in data story construction
- 3:31InReActable framework overview
- 5:41Pipeline and system design
- 7:06Prototype system demonstration
- 9:31User study and results
- 10:15Contributions and future work
- 11:42Q&A session
related talks
29mKeynote: LLMs, what are they good for?
20mEviza: A Natural Language Interface for Visual Analysis
38mLLMs for Self-serve Analytics - Lessons Learned
10mKeynote: Unlocking Corporate Data with Retrieval-Augmented Generation
24mInferencing underspecified natural language utterances in visual analysis
34mSearching for My Next Chart
26mThrough the looking glass w Data Rabbit: "A System built for Seeing"
30mWhat Can LLMs Do with Messy Data?