Paper Proposes Header-Only Framework for Table Quality Checks

A table with missing or unreliable cell values is normally treated as a dead end for automated knowledge graph work: there is nothing in the rows to classify. The paper submitted on 12 May 2026 by Marcelo Valentim Silva, Hannes Herrmann and Valerie Maxville argues the header row alone carries enough signal to keep going. Their system maps column headers to a set of interpretable FinalFormat types using curated lexical resources, while keeping a record of exactly which header tokens, called SourceKeywords, drove each assignment. That traceability is the paper’s central claim: a reviewer can see why a column was labelled “postal code” or “currency amount” rather than trusting a black box score.

Each assigned type then triggers validation rules drawn from a taxonomy of Data Quality Issues, flagging missing data, duplicate columns, domain violations, wrong data types and temporal mismatches. The authors roll these detections into a single unweighted metric they call HeadersIQ, meant as a quick read on how trustworthy a data source is before it reaches integration work, not a replacement for full downstream graph validation.

The numbers

The framework was tested against a mix of established benchmarks, including UCI, Prague, Kaggle, VizNet/Sato, SOTAB, T2Dv2 and the SemTab 2024 Metadata-to-KG track, together spanning around 120,000 header columns. Headers are sorted into 39 interpretable FinalFormat types, and the write-up itself runs 18 pages with four figures, presented at the Workshop on Quality of Knowledge Graphs at ESWC 2026 in Dubrovnik on 11 May 2026.

On the SemTab 2024 track specifically, the paper reports that strict ground-truth scoring came out modest. Rather than stop there, the authors ran a separate blinded diagnostic audit and found that many of the mismatches traced back to how the benchmark itself draws category boundaries, aliasing between near-synonym labels and ontology choices, not to implausible header readings from the system. They are careful to present that audit as evidence about disagreement patterns rather than a revised score, which keeps the strict benchmark result standing even while it argues the number understates the method in practice.

A parallel pathway in the framework also supports mapping results onto DBpedia and Wikidata, positioning the work as a preprocessing step that sits ahead of, rather than inside, the graph construction pipeline itself.