ReSpark: Leveraging Previous Data Reports as References to Generate New Reports with LLMs
summary
ReSpark uses LLMs to reverse-engineer hidden analysis workflows from past reports, modeling each report as a sequence of analysis segments (objective → transformation → insight), then adapts those segments to new datasets by generating adapted objectives, code, charts, and narrative. The implemented system produces draft code and visualizations with interactive refinement and insertion of segments. In evaluations versus GPT-4 baselines ReSpark’s reports score higher across quality metrics and the tool achieves a SUS score in top 10%.
AI-written summary · about
chapters (7)
- 0:00Introduction to report generation
- 1:47Leveraging prior reports as references
- 3:14Extracting analysis workflows from reports
- 6:48Reusing workflows for new data
- 7:54Demo of ResBock system
- 9:20Evaluation results and findings
- 10:25Conclusion and future work
related talks
46mSparking Success: Unveiling the Journey of Apache Spark Application Development
51mSpotlight on Emerging Data Science Worskspaces and Spark as Development Levers
01h 18mScalable Data Science with Spark and R
30mWhat Can LLMs Do with Messy Data?
43mIntroduction to Spark Datasets
59mSpark 2.0
38mThe Art of Compression: Crafting Insightful Summaries with LLMs
10mKeynote: Unlocking Corporate Data with Retrieval-Augmented Generation