Dango: A Mixed-Initiative Data Wrangling System using Large Language Model

TL;DR

Dango presents a mixed-initiative data-wrangling system that leverages demonstrations, natural-language prompts, and proactive clarification questions from LLMs to generate multi-table data-wrangling scripts. It extends a Domain Specific Language to support cross-table operations, and couples plan generation with step-by-step NL explanations and data-provenance visualization to improve interpretability and trust. In a 38-participant within-subject study, the CQ-enabled design (Condition C) reduced task time by up to 45% and lowered hallucinations, while boosting user confidence, with generalizability demonstrated on 24 additional tasks. The work shows that combining mixed-initiative interaction, transparent explanations, and provenance-aware feedback yields robust performance across diverse data-wrangling tasks, suggesting practical impact for data-cleaning pipelines.

Abstract

Data wrangling is a time-consuming and challenging task in a data science pipeline. While many tools have been proposed to automate or facilitate data wrangling, they often misinterpret user intent, especially in complex tasks. We propose Dango, a mixed-initiative multi-agent system for data wrangling. Compared to existing tools, Dango enhances user communication of intent by allowing users to demonstrate on multiple tables and use natural language prompts in a conversation interface, enabling users to clarify their intent by answering LLM-posed multiple-choice clarification questions, and providing multiple forms of feedback such as step-by-step natural language explanations and data provenance to help users evaluate the data wrangling scripts. We conducted a within-subjects user study with 38 participants and demonstrated that Dango's features can significantly improve intent clarification, accuracy, and efficiency in data wrangling. Furthermore, we demonstrated the generalizability of Dango by applying it to a broader set of data wrangling tasks.

Figures (8)

• Figure 1: Three core components: Intent Analysis, DSL-based Program Synthesis, and Program Evaluation & Refinement.
• Figure 2: User interface of Dango. In the table view (a), users can upload tables (b) or create new tables (c). Then, they can click the button (d) and start demonstrating their desired actions. Alternatively, they can express desired actions in natural language in a chatbox (e). Dango will interpret the demonstrations and/or NL descriptions in the backend and generate multiple-choice clarification questions when needed (f). Furthermore, to help users understand and validate the synthesized script, Dango explains it in NL step by step (g). Users can directly edit a step in natural language (h), delete a step (i), add a new step (j), save the script (k), remove the script (l), or regenerate the script (m). Users can click the button (n) to execute the script on copies of the original tables and verify its behavior without messing up the original demonstrations. Dango also renders a data provenance view to track the transformations performed on each table (o). Users can click table nodes, and the corresponding table content will appear in the table view.
• Figure 3: This figure shows a usage scenario of users using a chatroom to clarify their intent using NL prompts and answering multiple-choice clarification questions (left-hand side). Users can easily understand the program behavior by reading the step-by-step NL explanations. They can refine their program by directly editing the step-by-step NL explanations (right-hand side)
• Figure 4: User confidence on the data wrangling scripts when using conditions A, B, and C.
• Figure 5: User responses to the NASA TLX questionnaire (*: $p$-value < 0.05 based on ANOVA test).