ProvTales: A Dataset for Key Narrative Extraction from Provenance Graphs
Authors
Qi Jiang (School of Computer Science and Technology), Zhentao Zheng (Zhejiang University of Technology), Ziyu Huang (Zhejiang University of Technology), Binyan Gao (Zhejiang University of Technology), Xueqian Zheng (Zhejiang University of Technology), Xuhui Ma (Zhejiang University of Technology), Ronghua Liang (Zhejiang University of Science and Technology), Guodao Sun (Zhejiang University of Technology)
Presentation
- Session
- Story time
- Time
- Thursday, Nov 12, 08:00 – 08:12 (US/Eastern) · session 08:00 – 09:30
- Location
- Hall America north
Keywords
Visualization dataset, Provenance graph, Visual narrative, Reverse synthesis
Abstract
Visual analytics produces provenance graphs that capture analytic states and their semantic relationships. To communicate findings derived from such exploration, data narratives have proven effective by organizing key insights into coherent sequences accompanied by visualizations. However, to our knowledge, this task still lacks a benchmark dataset, hindering controlled comparison and systematic evaluation of existing methods. To bridge this gap, we present ProvTales, the first large-scale dataset for converting visual exploration histories into data narratives. ProvTales follows a narrative-first, provenance graph-second construction pipeline. Specifically, we first use large language models (LLMs) to generate narratives progressively through topic planning, blueprint design, and data-fact instantiation. Next, we treat each narrative as a backbone and expand it into a provenance graph by introducing auxiliary non-narrative nodes that emulate realistic exploratory behaviors. The ProvTales further defines constrained and unconstrained settings based on whether an explicit communication intent is available. Comprehensive statistical analyses and evaluations demonstrate the quality and diversity of ProvTales. Finally, we discuss limitations in data construction and possible directions for future improvement. The dataset is publicly available at https://huggingface.co/datasets/ZtZheng/ProvTales.
For Practitioners
This paper is particularly relevant to visual analytics and visualization developers, data scientists, data journalists, BI analysts, and engineers building AI/LLM-based analytical assistants. In practice, they can apply its ideas to summarize complex analysis histories, identify the most important findings and visualizations, remove irrelevant exploratory branches, generate clearer reports or data stories, and recommend useful next analysis steps.