Authors: Anna-Christina Glock, Thomas Hütter, Johannes Fürnkranz, Wolfram Wöß, Christine Dominka-Kiss, Lisa Ehrlinger
Organizations: Software Competence Center Hagenberg, Austria · Johannes Kepler University Linz, Austria · Austrian Post Vienna, Austria · Hasso Plattner Institute, University of Potsdam Potsdam, Germany
The validation of data, such as customer and employee data, is an important task in many organizations. Errors in data can have severe consequences. For example, a wrong drug unit in a patient record can lead to life-threatening medication errors, and a missing street number in an address to failed deliveries. Companies often employ rule-based enterprise data quality (DQ) tools, which allow domain experts to specify rules to validate the data over time. While rule-based DQ tools are computationally efficient and provide explainable reports, maintaining a comprehensive rule set manually is challenging, as domain experts often overlook essential rules, especially in complex domains and large data volumes. Hence, closing these gaps remains an open problem in practice. In this paper, we address the challenge of automated DQ rule generation. For this, we formalize a generalizable generate-filter framework and introduce LeDQeR, an LLM-based DQ rule generation approach. First, a large language model (LLM) generates candidates rules from an observed dirty data tuple for a given rule-based DQ tool syntax. Second, we apply four filter techniques that ensure the (i) executability, (ii) correctness, and (iii) generalizability, and avoid (iv) redundancy of the generated rules. An extensive experimental evaluation suggests that LeDQeR is able to produce effective and compact rule sets for various datasets and error types.
Figures & tables
Figure 1. LeDQeR extends the rule set of a rule-based enterprise DQ tool. Incoming data is validated against an existing rule set R={r1,…,rm} . Errors covered by a rule are detected, while errors not covered by any rule pass the validation unnoticed and cause failures in downstream processes. LeDQeR closes these gaps: undetected dirty tuples are passed to the generate-filter pipeline, which produces new rules {rm+1,rm+2} that extend the existing rule set such that the same error type is detected in the future.
ID
Street
FirstName
PostalCode
City
1
Unknown
Ina-Maria
A-1220
Wien
2
Brahmsplatz
Hans
1040
Wien, Wieden
3
Berggasse
Anna
A-4020
Linz
4
Feldgasse
Anna-Maria
A6020
Innsbruck
5
Unknown
Maria
8010
Graz
Type of error: ■ Embedded Value
Table 1. Subset of the PCI dataset as running example.
Figure 2. Generate-filter framework for DQ rule generation.
Figure 3. Architecture of LeDQeR : an LLM generates candidate rules C , which are progressively refined through four filters C⊇C′⊇C′′⊇C′′′ until a final rule set R is returned.
Dataset
# Columns
# Dirty Tuples
# Error Types
Beers
10
2,410
1
Flights
7
1,904
3
Hospital
19
407
1
Tax
15
6,000
2
PCI
14
6,000
9
Table 2. Dataset characteristics used in our experiments.
GLM-4.7
Gemma-4
Qwen3-Coder
no Parameter
Parameter
no Parameter
Parameter
no Parameter
Parameter
Dirty & Clean
Beers
No. Rules
2,534
2,569
2,984
3,119
3,223
3,137
% Executable
0.28
0.84
0.99
1.00
0.98
0.94
Flights
No. Rules
2,002
2,748
3,154
902
7,860
7,804
% Executable
0.45
0.77
0.95
0.93
0.98
0.94
Table 3. Syntactic filter: Number of generated rules and the corresponding percentage of executable rules across three LLMs, five D , and three input variants (Dirty & Clean, Dirty & Type, Just Dirty), and two S variations (with vs. without parameter information).
Figure 4. Correctness filter: Aggregated F1-Score per error type across the three LLMs, three input variants, and five datasets.
Figure 5. Coverage filter: Aggregated recall per error type across the three LLMs, three input variants, and five datasets.
Figure 6. Precision and recall of each rule measured on an unseen dataset Dtest for each of the three LLMs, three data input variants, and five datasets. The results have been ordered by F1 score.
Figure 7. Redundancy filter: Percentage of covered clean and dirty tuples found by the rules in R , calculated for each of the five Dtest , three input variants and three LLMs with different precision settings for the rules in C′′′ .
Step
Xie et al.
Schneider
LeDQeR
Input
Requirements + Metadata
Template + Metadata
Dirty tuples
Generation
One-time
One-time
Incremental
Filtering
Syntactic
✓
✓ a
✓
Correctness
✗
✓ a
✓
Coverage
✗
✗
✓
Table 4. Comparison of LeDQeR with related LLM-based rule generation frameworks. Xie et al. ( Xie et al., 2025 ) and Schneider ( Schneider, 2025 ) .
Oct 7, 2026·Philipp Jung, Katinka Becker, Felix Biessmann +4
Berliner Hochschule für Technik Berlin, Germany · Physikalisch-Technische Bundesanstalt Berlin, Germany · Einstein Center Digital Future Berlin, Germany +3