PaperlayerView exact revision on arXiv

Source-linked AI summary

An Explainable Header-Centric Framework for Large-Scale Semantic Table Interpretation and Data Quality Assessment

Marcelo Valentim Silva, Hannes Herrmann, Valerie Maxville

arXiv:2610.10541v1cs.AIcs.CL

TL;DR

Poorly annotated tables make semantic integration difficult, and metadata-only settings leave headers as key evidence when cell values are unavailable or unsuitable. The paper uses header-centric typing and traceable quality checks across benchmarks, showing large-scale practical support while diagnosing disagreement in KG-oriented evaluation.

  • Problem

    Poorly annotated tables hinder semantic integration, and automated profiling lacks the semantic interpretation needed to determine which attribute constraints apply.

    Source: §1 · p. 1, §2.2 · p. 4

  • Method

    The framework maps headers to FinalFormat types, preserves token-level justifications through SourceKeywords, and aggregates activated quality checks into HeadersIQ.

    Source: §1.2 · p. 3, §6 · p. 15

  • Results

    Across around 120,000 header columns from seven benchmarks, the framework supports large-scale metadata-only interpretation and provides diagnostic evidence on SemTab 2024 KG-oriented evaluation.

    Source: §6 · p. 15

  • Takeaways & Limitations

    The framework is presented as a practical contribution to metadata-driven semantic annotation, quality-aware tabular preparation, and trustworthy knowledge-graph integration.

    Source: §6 · p. 15

    The approach depends on meaningful headers and struggles with long-tail terms; several reported results measure coverage rather than direct correctness.

    Source: §5.5 · p. 15

Abstract

from arXiv · show

Knowledge Graph (KG) quality depends not only on downstream graph validation, but also on the quality of tabular metadata used before integration. In metadata-only Semantic Table Interpretation (STI), where cell values are unavailable, noisy, or unsuitable, column headers become a critical source of semantic evidence for traceable KG preparation. We present an explainable, header-centric framework for metadata-only Column Type Annotation (CTA) and Data Quality Assessment (DQA). The framework maps headers to 39 interpretable FinalFormat types using curated lexical resources and preserves token-level traceability through SourceKeywords. Each assigned type activates validation rules based on a taxonomy of Data Quality Issues (DQIs), producing detections such as missing data, duplicates, domain violations, wrong data type, and temporal mismatch. These detections are aggregated into HeadersIQ, a lightweight, unweighted data source-level quality metric. The framework was evaluated across heterogeneous benchmarks, including UCI, Prague, Kaggle, VizNet/Sato, SOTAB, T2Dv2, and the SemTab 2024 Metadata-to-KG track, comprising around 120,000 header columns. The results show broad practical coverage across noisy real-world metadata, while a parallel KG-mapping pathway supports alignment to DBpedia and Schema.org. On the SemTab 2024 Metadata-to-KG track, the official GT-strict evaluation was modest. However, a blinded diagnostic audit indicates that many mismatches reflect benchmark granularity, aliasing, and ontology-selection effects rather than wholly implausible header-centric predictions. We report this audit as diagnostic evidence on disagreement patterns, not as revised benchmark performance. Overall, the paper presents a reusable workflow for metadata-driven semantic annotation, data source-level quality monitoring, and KG-oriented benchmark diagnosis.

1. Introduction

The introduction motivates header-centric semantic typing as a metadata-only route to more interpretable tabular integration and quality assessment. It presents a workflow connecting header-derived types, traceability, and data quality checks, alongside a separate KG-mapping pathway.

1. Introduction

  • Poorly annotated tables make semantic integration difficult; column type annotation assigns meaningful types that support alignment and downstream knowledge-graph integration.The introduction frames this problem amid expanding open-data repositories and domain-specific tabular sources.Source: §1 · p. 1

1.1. Historical context and renewed interest

  • Sherlock (2019) cautioned against header-centric matching, while Liu et al. (2023) report growing use of header signals in competitive CTA systems.SemTab 2024 introduced a metadata-only track, and AdaTyper (2023) [13] prioritises header analysis before cell-based fallback.Source: §1.1 · p. 2
  • Dynamic dictionary expansion, normalization, and abbreviation mappings target limited type coverage and noisy headers, while context-dependent ambiguity remains challenging.The FinalFormat layer extends semantic coverage beyond primitive categories.Source: §1.1 · p. 2

1.2. Approach and Contributions

  • The framework uses headers as primary semantic evidence when cell values are noisy, unavailable, sensitive, or unsuitable, then branches into type detection and KG mapping.Type assignments activate targeted checks grounded in a Data Quality Issues taxonomy; the parallel pathway aligns headers to DBpedia and Schema.org.Source: §1.2 · p. 2
  • FinalFormat types use curated lexical resources and token-level traceability, while HeadersIQ aggregates detected violations for data source-level quality monitoring.SourceKeywords identify the matched header tokens supporting each assignment.Source: §1.2 · p. 3
  • The paper contributes a broad benchmark study and proposes HeadersIQ and benchmark diagnosis as parts of its quality-monitoring contribution.A correctness-oriented Statistical Audit is described as ongoing work, beyond the reported coverage-focused indicators.Source: §1.2 · p. 3

2. Related Work

The related work surveys CTA paradigms and identifies header-only annotation as an underexplored setting. It then connects semantic interpretation to attribute-level validation and grounds concrete checks in a Data Quality Issues taxonomy.

2.1. Column Type Annotation (CTA) Paradigms

  • CTA spans dictionary matching, iterative constraint propagation, feature-based learning, and deep or embedding-based systems; relatively few systems use headers, and fewer avoid cell values.Examples include Sherlock, Sato, TURL, and Doduo.Source: §2.1 · p. 3

2.2. Attribute-Level Data Quality Dimensions

  • Semantic type detection supplies the link between attribute labels and dimension-aware validation, building on prior work connecting type assignments to executable rules.The cited dimensions include accuracy, completeness, consistency, uniqueness, and timeliness.Source: §2.2 · p. 4

2.3. Data Quality Issues (DQIs) and Taxonomic Grounding

  • The framework adopts a 22-issue DQI taxonomy to map concrete error patterns to quality dimensions, including missing data, duplicates, temporal mismatch, and domain violations.Cross-table referential-integrity checks fall outside scope; the study focuses on issues deterministically activated through attribute-level interpretation.Source: §2.3 · p. 4

3. Methodology

The methodology preprocesses column metadata, then branches into FinalFormat typing for DQA and KG mapping; type-triggered checks and NIL review support quality monitoring and dictionary curation.

3.1. Data Ingestion

  • Ingestion extracts headers, descriptions, and key indicators from varied tabular formats; curated dictionaries grew to roughly 2,800 format entries and over 1,000 abbreviation mappings supporting 39 types.Coverage results reflect the dictionary snapshot available when each benchmark was processed.Source: §3.1 · pp. 5–6

3.2. Column Header Preprocessing

  • Header preprocessing normalizes tokens, expands abbreviations and variants, and records matched SourceKeywords to make each FinalFormat assignment traceable.The SourceKeyword is the dictionary-matched trigger token from the cleaned header or optional description.Source: §3.2 · pp. 6–7

3.3. Semantic Type Detection

  • FinalFormat defines 39 interpretable types between atomic data types and ontology classes, assigned through exact, substring, and fuzzy dictionary matching.SourceKeywords preserve the supporting header or description tokens for explanation and auditing.Source: §3.3 · p. 7

3.4. Knowledge Graph Mapping

  • KG mapping aligns preprocessed headers to DBpedia and Schema.org, while vocabulary coverage and modeling choices can constrain mapping accuracy without affecting FinalFormat-based DQA.Examples include accurate mappings and NIL or overly specific results.Source: §3.4 · p. 8

3.5. Data Quality Assessment

  • Each inferred FinalFormat activates predefined validation functions grounded in the DQI taxonomy, producing cell-level violation records.Checks include type-specific constraints such as numerical bounds and date structure; numerical intervals are configurable by FinalFormat.Source: §3.5 · pp. 8–9

3.5.1. HeadersIQ: A Metric for Data/Information Quality Monitoring

  • HeadersIQ is an unweighted source-level metric representing the proportion of evaluated cells with no detected violations under activated rules.It is normalized to a 0–100 score.Source: §3.5.1 · p. 9

3.6. Feedback, NIL Analysis & Knowledge Curation

  • NIL headers enter a feedback cycle that ranks unmatched tokens, filters unstable entries, manually reviews frequent candidates, and adds only validated, interpretable terms to the dictionaries.Source: §3.6 · p. 9

4. Results and Coverage Analysis

The evaluation finds broad FinalFormat coverage across heterogeneous benchmarks, with 15 core types covering most Kaggle and VizNet headers while additional types support finer-grained checks.

4.1. Data Sources and Benchmarking Scope

  • Across seven heterogeneous benchmarks, the evaluation examines FinalFormat assignments, SourceKeywords, and DBpedia/Schema.org links; Universe B contains 118,639 unique Kaggle and VizNet headers.The benchmarks include structured repositories, web-table collections, user-generated tables, and KG-grounded datasets.Source: §4.1 · p. 10, §4 · p. 10, §4.2 · p. 10

4.2. Semantic Type Detection: Coverage and Diversity

  • Kaggle covers all 39 FinalFormat types, while VizNet and Kaggle have higher SourceKeywords counts due to their scale.The framework reports broad lexical coverage and assignment robustness across heterogeneous, noisy repositories including Kaggle and VizNet.Source: §4.2 · p. 10, §4 · p. 10
  • 15 core FinalFormats deterministically cover the 76-label Sherlock/Sato space, and together they cover approximately 80% of more than 118,000 Kaggle and VizNet headers.The 15 types preserve distinctions relevant to data quality while merging redundant semantic variants into compact categories.Source: §4.2 · pp. 10–11
  • 24 additional FinalFormats account for approximately 6% of Universe B, exceeding 7,000 headers and enabling finer-grained DQI checks beyond the Sherlock/Sato taxonomy.These include bounded numerical and domain-specific formats such as percentage, chemical, health, and geographic types.Source: §4.2 · p. 11
  • Approximately 13% of Universe B headers are NIL, meaning no FinalFormat matched after normalization and abbreviation expansion; the remaining 87% are Valid.Source: §4.2 · p. 11
  • An approximately 80% increase in dictionary entries, from about 2,100 to 3,800, supported approximately 103 times the practical coverage, from about 1,000 to 103,000 valid assignments.The scalability comparison is reported in Table 5.Source: §4.2 · p. 11, Table 5 · p. 11

4.3. NIL Token Improvement Process

The NIL review finds that unresolved headers are mostly low-information tokens, while stable, interpretable domain-specific abbreviations are added to the mappings.

4.3. NIL Token Improvement Process

  • Most of the top 100 NIL cases are extremely short codes or symbols with little or no semantic information, so they fall outside meaningful FinalFormat mappings.The authors characterize the remaining NIL proportion mainly as low-information tokens or highly specialized terminology.Source: §4.3 · p. 11
  • Domain-specific or plausible abbreviations are added only when their meaning is stable and the resulting mapping remains interpretable.Source: §4.3 · p. 11

4.4. Validity, NIL Rates, and Robustness

Official SemTab 2024 scores are modest, while a separate blinded audit identifies disagreement patterns; coverage and ambiguous-header results also show strengths and boundaries across benchmarks.

4.4. Validity, NIL Rates, and Robustness

  • FinalFormat has the highest Valid percentages across benchmarks, including 100% Valid on Sato; KG mapping coverage is notably lower for UCI and Kaggle.Sato also reaches 100% Valid for DBpedia, while UCI and Kaggle have heterogeneous headers and cases without suitable ontology matches.Source: §4.4 · p. 12

4.5. Handling Ambiguous Headers

  • Normalization, abbreviation expansion, and SourceKeywords help resolve compressed or ambiguous headers, including binary cues, geographic terms, and medical abbreviations.The examples include “has,” “birth place,” and “trestbps” mapped to bloodpressure; the authors describe the resulting FinalFormats as interpretable and rule-aligned.Source: §4.5 · p. 12

4.6. SemTab 2024 Metadata-to-KG Evaluation

  • 63/141 columns achieve Hit@1 = 0.45 and 67/141 achieve Hit@5 = 0.48 under the official SemTab 2024 GT-strict evaluation.The head-to-head figures use the organizers’ ground truth and scripts.Source: Table 7 · p. 13, §4.6 · p. 13
  • The blinded audit classifies 43/141 cases as GT-Refinement and 6/141 as ontology-consistent alternatives; cumulative diagnostic support rises from 0.45 to approximately 0.79, not as revised benchmark scores.The audit attributes a substantial subset of strict-GT mismatches to granularity, aliasing, or ontology-selection effects.Source: §4.6 · p. 13, Table 7 · p. 13

4.7. Illustrative Case Studies from Real Datasets

Real-dataset examples show how inferred column semantics activate checks for missingness, type, uniqueness, and domain violations, while an age column passes its expected-range checks.

4.7. Illustrative Case Studies from Real Datasets

  • The Online Retail Description column, classified as String, has 1,454 blank or null values and one numeric value in the textual field.Source: §4.7 · p. 13
  • CustomerID, classified as IDcolumn, has 135,080 missing values, 541,830 repeated identifiers, and 537,536 explicit uniqueness violations.The example illustrates how identifier semantics activate completeness and uniqueness constraints.Source: §4.7 · p. 14
  • The heart-rate column thalach, classified as heartrate, triggers a domain violation because a value exceeds the medically plausible upper bound.Source: §4.7 · p. 14
  • The age column, classified as age, passes validation with observed values from 29 to 77 within its expected range.Source: §4.7 · p. 14

4.8. HeadersIQ: Data Source-Level Data Quality Measurement

HeadersIQ summarizes detected data-quality violations at the data-source level, with scores normalized for source size. The metric supports comparison and monitoring across heterogeneous tabular sources.

4.8. HeadersIQ: Data Source-Level Data Quality Measurement

  • Dataset 45 has 7 cells with at least one detected violation among 4,242 cells.The passage identifies the dataset as having 303 rows and 14 columns.Source: §4.8 · p. 14
  • Scores near 100 indicate few detected violations; lower scores reflect more structural, semantic, or representational issues.Normalizing by data-source size enables comparisons across heterogeneous sources.Source: §4.8 · p. 14
  • HeadersIQ supports ranking and monitoring tabular sources before knowledge-graph integration as a lightweight governance signal.The passage relates this use to SemTab 2024 IsGold? and KG2Tables-style data-source reliability questions.Source: §4.8 · p. 14

5. Discussion

The discussion presents header-centric typing as practically effective under metadata-only constraints, with a compact type vocabulary that scales across heterogeneous sources. It also distinguishes diagnostic benchmark evidence from direct correctness measures and identifies meaningful headers and long-tail terms as limitations.

5.1. Reassessing Header-Centric Type Detection

  • Headers alone can support precise, scalable type assignments after preprocessing, dictionary enrichment, and abbreviation expansion.The 39 FinalFormat types bridge primitive types and fine-grained ontology classes, while SourceKeywords retain finer lexical evidence for auditing and KG interpretation.Source: §5.1 · p. 14
  • The framework targets metadata-only interpretation, not outperforming systems designed for richer supervision or table-value context on their native task.Its objective is practical header-centric typing at scale with explicit traceability and DQA activation.Source: §5.1 · p. 14

5.2. General-Purpose Core and Domain-Specific Long-Tail Modelling

  • An approximately 80% dictionary expansion enabled more than 100-fold coverage growth, from about 1,000 UCI assignments to roughly 103,000 across over 118,000 headers.The 39 FinalFormats comprise 15 core categories and 24 domain-specific types, with consistent results across UCI, Prague, Sato/VizNet, Kaggle, SOTAB, and T2Dv2.Source: §5.2 · p. 15

5.3. Data Source-Level Quality via HeadersIQ

  • HeadersIQ provides scalable data-quality monitoring across sources of varying sizes and is relevant to data-source reliability questions.The discussion connects this use to SemTab 2024 IsGold? and KG2Tables-style work.Source: §5.3 · p. 15

5.4. Interpreting Ground Truth in SemTab 2024

  • SemTab 2024 yielded Hit@1 = 0.45 and Hit@5 = 0.48; a blinded audit linked many GT-strict mismatches to granularity, aliasing, or ontology selection.The audit is diagnostic evidence, motivating multi-label or graded evaluation and explicit human-in-the-loop validation rather than revised benchmark performance.Source: §5.4 · p. 15

5.5. Limitations and Future Work

  • The framework depends on meaningful headers and has difficulty with long-tail terms; several results measure coverage rather than direct correctness.Ongoing work includes a Statistical Audit of FinalFormat correctness and KG vocabulary-gap analysis; proposed directions include combining header and value signals, multilingual models, and LLM-based disambiguation.Source: §5.5 · p. 15

6. Conclusion

The conclusion presents an explainable, header-centric framework for metadata-only semantic typing and data-quality assessment. Evaluation across about 120,000 columns and a diagnostic SemTab audit supports its use for large-scale interpretation and benchmark diagnosis.

6. Conclusion

  • The framework assigns 39 FinalFormat types using normalization, curated dictionaries, and SourceKeywords, with mappings to DBpedia and Schema.org and data-quality checks aggregated in HeadersIQ.Source: §6 · p. 15
  • Experiments across around 120,000 columns from seven benchmarks and a diagnostic SemTab 2024 audit support large-scale metadata-only interpretation and KG-oriented benchmark diagnosis.The conclusion positions the framework for metadata-driven annotation and quality-aware tabular preparation.Source: §6 · p. 15

Declaration on Generative AI

The authors report using ChatGPT/GPT5 for language and formatting assistance and for code development, debugging, and explanation; they reviewed and edited the content and take responsibility for it.

Declaration on Generative AI

  • ChatGPT/GPT5 assisted with writing and code tasks, after which the authors reviewed and edited the work and accepted responsibility for its content.Source: Declaration on Generative AI · p. 16
Loading 2610.10541v1…