Reconciling Conflicting Data Curation Actions: Transparency Through Argumentation

TL;DR

The paper tackles conflicts in collaborative data cleaning by modeling competing actions as an argumentation framework and translating the AF to a logic program $P_{AF}$ to compute transparent solutions using well-founded (grounded) and stable semantics. This yields a unique, 3-valued verdict (accepted, rejected, undecided) and, when needed, enumerates multiple merged recipes through stable extensions, supporting user-driven reconciliation. A detailed running example with OpenRefine operations demonstrates how conflicting edits can be resolved and justified formally, leading to a merged, reproducible data-cleaning recipe. The work contributes a principled, provenance-aware approach and ongoing tooling to integrate argumentation-based reconciliation into practical data-cleaning workflows like OpenRefine.

Abstract

We propose a new approach for modeling and reconciling conflicting data cleaning actions. Such conflicts arise naturally in collaborative data curation settings where multiple experts work independently and then aim to put their efforts together to improve and accelerate data cleaning. The key idea of our approach is to model conflicting updates as a formal \emph{argumentation framework}(AF). Such argumentation frameworks can be automatically analyzed and solved by translating them to a logic program $P_{AF}$ whose declarative semantics yield a transparent solution with many desirable properties, e.g., uncontroversial updates are accepted, unjustified ones are rejected, and the remaining ambiguities are exposed and presented to users for further analysis. After motivating the problem, we introduce our approach and illustrate it with a detailed running example introducing both well-founded and stable semantics to help understand the AF solutions. We have begun to develop open source tools and Jupyter notebooks that demonstrate the practicality of our approach. In future work we plan to develop a toolkit for conflict resolution that can be used in conjunction with OpenRefine, a popular interactive data cleaning tool.

Figures (4)

• Figure 1: (a) AF with four arguments $\mathsf{a}$, $\mathsf{b}$, $\mathsf{c}$, $\mathsf{d}$ and their attack relation. (b) The unique, 3-valued grounded solution: $\mathsf{a}$ is accepted ($\textcolor{RoyalBlue}{\mathsf{blue}}$), $\mathsf{b}$ is defeated ($\textcolor{DarkOrange}{\mathsf{orange}}$), and $\mathsf{c}$, $\mathsf{d}$ are undecided ($\textcolor{DarkYellow}{\mathsf{yellow}}$). $G_\mathsf{AF}$ has two stable solutions: The undecided argument $\mathsf{c}$ can be chosen as accepted and $\mathsf{d}$ as defeated, as in (c), or vice versa as in (d), yielding two separate stable solutions.
• Figure 2: Individual attack relations and visualized attack graph (with recipe execution order)
• Figure 3: The grounded extension of Figure \ref{['fig:atk-graph']} where ($\textcolor{RoyalBlue}{\mathsf{blue}}$) actions are accepted, ($\textcolor{DarkOrange}{\mathsf{orange}}$) actions are rejected, and ($\textcolor{DarkYellow}{\mathsf{yellow}}$) actions are undecided.
• Figure 4: A stable extension of Figure \ref{['fig:wf_dc']} with actions appearing in $\textcolor{lightblue}{\mathsf{light~blue}}$ being additionally accepted and those appearing $\textcolor{lightorange}{\mathsf{light ~orange}}$ additionally being rejected.