PaperlayerView exact revision on arXiv

Source-linked AI summary

GitTables: A Large-Scale Corpus of Relational Tables

Madelon Hulsebos, Çağatay Demiralp, Paul Groth

arXiv:2106.07258v5cs.DBcs.LG

TL;DR

Web-derived table corpora provide a limited representation of database-like tables, constraining resources for research beyond the Web. GitTables contributes a large corpus of GitHub tables with semantic column annotations. Its tables differ from Web tables in dimensions, content, and semantic-type distribution, and the corpus supports several data-management applications.

  • Problem

    Existing Web-table corpora poorly represent database-like tables, while models based on them have limited applications beyond the Web.

    Source: §1 · p. 2

  • Method

    GitTables is a corpus of 1M GitHub CSV tables annotated with semantic types and related metadata from DBpedia and Schema.org.

    Source: §1 · p. 2

  • Results

    GitTables has significantly larger dimensions than Web tables, and its content and semantic-type distribution differ; a model accurately distinguishes columns by corpus.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The corpus is used for semantic type detection, schema completion, and benchmark datasets for typical data-management tasks.

    Source: §1 · p. 3

    The design targets tables considerably larger and more heterogeneous than typical Web tables, with corpus scale intended to exceed open-data portals.

    Source: §3.1 · p. 5

Abstract

from arXiv · show

The success of deep learning has sparked interest in improving relational table tasks, like data preparation and search, with table representation models trained on large table corpora. Existing table corpora primarily contain tables extracted from HTML pages, limiting the capability to represent offline database tables. To train and evaluate high-capacity models for applications beyond the Web, we need resources with tables that resemble relational database tables. Here we introduce GitTables, a corpus of 1M relational tables extracted from GitHub. Our continuing curation aims at growing the corpus to at least 10M tables. Analyses of GitTables show that its structure, content, and topical coverage differ significantly from existing table corpora. We annotate table columns in GitTables with semantic types, hierarchical relations and descriptions from Schema.org and DBpedia. The evaluation of our annotation pipeline on the T2Dv2 benchmark illustrates that our approach provides results on par with human annotations. We present three applications of GitTables, demonstrating its value for learned semantic type detection models, schema completion methods, and benchmarks for table-to-KG matching, data search, and preparation. We make the corpus and code available at https://gittables.github.io.

1 INTRODUCTION

Existing table corpora are dominated by Web tables, which poorly represent database-like tables and limit applications beyond the Web. GitTables contributes a large corpus of GitHub CSV tables with semantic column annotations, and its structure and topical coverage differ from Web-table corpora.

1 INTRODUCTION

  • Web-table corpora skew representation away from database tables, limiting the reach of models based on them beyond the Web.For example, “ID” is not among WDC WebTables’ twenty most frequent column names, and its tables are significantly smaller than typical database tables.Source: §1 · p. 2
  • GitTables provides 1M relational tables extracted from GitHub CSV files, with continued curation aimed at at least 10M tables.The corpus is intended to complement existing collections and support deep transfer learning for relational data.Source: §1 · p. 2
  • GitTables columns receive semantic types, expected atomic data types, descriptions, and hierarchical type relations from DBpedia and Schema.org.These annotations support applications such as recommending visualizations or imputing missing values from related columns.Source: §1 · p. 2
  • GitTables has significantly larger row and column dimensions, and its content and semantic-type distribution differ from Web tables.A model detecting data shift accurately distinguishes columns from the two corpora.Source: §1 · p. 2
  • Annotation confidence scores allow users to filter labels for their intended use case.The example scores are 1 for the Species column’s species type and 0.7 for the Organism Group column’s element group type.Source: Figure 2 · p. 3
  • The paper presents GitTables applications in semantic type detection, schema completion, and benchmark datasets for data-management tasks.The semantic type detection model achieves high prediction accuracy.Source: §1 · p. 3

2 RELATED WORK

Prior table corpora come from varied sources and serve different tasks, but many remain small or Web-focused. Annotated datasets add resources for semantic and knowledge-graph tasks, with GitTables positioned as a larger corpus of relational tables and types.

2.1 Large-scale table corpora

  • WDC WebTables and Dresden Web Table Corpus contain 59M–90M HTML-derived tables, but their small dimensions and different content limit generalization.The section contrasts them with database-like tables.Source: §2.1 · p. 3
  • WikiTables provides approximately 2M Wikipedia tables suited to question answering, but its tables are as small as those in WDC WebTables.Source: §2.1 · p. 3
  • Open Data Portal Watch analyzed 108K parsed tables from 227K CSV files, while VizNet combines 31M tables from WebTables, ManyEyes, Plotly, and Open Data portals.Analyses of VizNet suggest that non-Web tables have different internal structures.Source: §2.1 · p. 4

2.2 Table datasets with column annotations

  • T2Dv2 manually annotates WDC WebTables with DBpedia correspondences, while SemTab benchmarks KG matching using tables largely from Wikidata and Wikipedia.The T2Dv2 target was found to be trivial for KB matching because many entities were obviously linkable; GitTables is proposed for future SemTab versions.Source: §2.2 · p. 4
  • TURL uses WikiTables annotated with 255 Freebase semantic types to fine-tune pretrained table representations for downstream tasks.Table 2 characterizes annotated relational table datasets and notes that GitTables has over 2K semantic types.Source: §2.2 · p. 4, Table 2 · p. 4

3 GITTABLES

GitTables is designed as a large, diverse corpus of database-like CSV tables, enriched with column semantics for data-management research. Its construction combines GitHub search, parsing and curation, and column annotation from DBpedia and Schema.org.

3.1 Design principles of GitTables

  • The design criteria call for scale beyond Open Data portals, larger and more heterogeneous database-like tables, broad topical coverage, and semantic annotations.Source: §3.1 · p. 5
  • GitHub search returned 92,191,141 CSV files, and the authors use 67K English WordNet nouns as diverse query topics.CSV is selected for its wide use in storing raw structured data; WordNet topics introduce some bias toward English tables.Source: §3.1 · pp. 5–6

3.2 Extracting CSV files from GitHub

  • Because GitHub limits each search response to 1,000 files, the extraction pipeline segments topic queries by file size and collects paginated results.Files larger than 438 kB cannot be retrieved; forked-repository results are excluded to minimize duplication.Source: §3.2 · p. 6

3.3 Parsing and curating tables from CSV files

  • CSV curation parses files with Pandas and filters tables by licensing, minimum size, header quality, and social-media column names.The pipeline also anonymizes likely personal data using fake values; 0.3% of GitTables columns contain fake values.Source: §3.3 · p. 7

3.4 Annotating tables with column semantics

  • Columns receive semantic types from DBpedia and Schema.org, with hierarchical relations, domains, descriptions, and other metadata where available.The syntactic method matches preprocessed column names directly to ontology types; the semantic method uses FastText embeddings and discards annotations with very low similarity scores.Source: §3.4 · pp. 7–8

4 ANALYSIS

The analysis examines GitTables 1M across 97 topic subsets, finding that its dimensions, content, and topical coverage differ from Web-table corpora. It also evaluates annotation coverage and quality, using T2Dv2 as a proxy because GitTables lacks semantic-type ground truth.

4.1 Corpus statistics

  • The 1,021,143-table analysis set averages 142 rows and 12 columns, compared with 15 rows and 5 columns for Web-table corpora; nearly 58% of columns are numeric.GitTables tables also show long-tailed row and column counts.Source: §4.1 · p. 9
  • 75% of source repositories contribute at most five tables, while a few repositories supply many tables that may be joined or unioned into larger tables.Manual inspection found that the high-contribution repositories contain snapshots of the same or similar databases.Source: §4.1 · p. 9
  • Semantic annotation covers 71% of table columns on average, versus 26% for syntactic annotation; semantic annotation also covers more tables and unique types.Across the two ontologies, semantic annotation covered 960K tables and 2.4K unique types on average, versus 730K tables and 756 types for syntactic annotation.Source: §4.1 · p. 9, Figure 4 · p. 10, Table 5 · p. 10
  • Semantic annotation similarity scores peak near 1, reflecting syntactic resemblance, while the remaining distribution centers around 0.75; users can set a threshold to select annotations.The scores are attached to semantic annotations as confidence indicators.Source: §4.1 · p. 10, Figure 4 · p. 10

4.2 Corpus content

  • A domain classifier distinguished GitTables from VizNet columns with 93% (±0.04) accuracy, indicating different data distributions and complementary corpus value.The classifier used 5,000 deduplicated columns from each corpus and 10-fold cross-validation.Source: §4.2 · p. 10
  • GitTables’ top semantic types include id, type, value, and min, unlike WebTables’ list of name, date, title, artist, description, size, type, location, model, and year.The authors identify the frequent database-like type id as evidence of GitTables’ distinct topical coverage.Source: §4.2 · p. 11, Figure 5 · p. 11
  • GitTables’ geographic columns make up slightly over 1% and primarily represent English-speaking countries and cities; its few population-related columns skew toward Western countries.The profile uses values associated with relevant Schema.org semantic types.Source: §4.2 · p. 11, Table 6 · p. 12

4.3 Annotation quality

  • In manual review, the semantic approach yielded better annotations in 63 (± 14) of 148 cases, versus 37 (± 3) where T2Dv2 was clearly better; the authors judged both approaches suitable for training and evaluating DL models.The semantic approach matched T2Dv2 on 54% of evaluated columns, while the syntactic approach matched it on 61%.Source: §4.3 · p. 12

5 APPLICATIONS

The applications evaluate GitTables for semantic column type detection, schema completion, and data-management benchmarks. Results show strong detection performance on GitTables, relevant schema suggestions, and challenging table-to-KG matching and search use cases.

5.1 Semantic column type detection

  • A GitTables-trained semantic type detector achieves macro F1 0.86 (± 0.02), while a VizNet-trained model scores 0.66 (± 0.03) on GitTables columns.On VizNet samples, the VizNet-trained model scores macro F1 0.77 (± 0.02), indicating its weaker cross-corpus performance on GitTables.Source: §5.1 · p. 13, Table 7 · p. 13

5.2 Schema completion

  • Suggested schema completions are relevant to their target prefixes, with an average whole-schema cosine similarity of 0.49 on a [-1,1] scale.The suggestions are selected by comparing Universal Sentence Encoder embeddings of target and GitTables schema prefixes.Source: §5.2 · pp. 13–14, Table 8 · p. 14

5.3 Data management benchmarks

  • The 1,101-table semantic-type benchmark exposes challenges for table-to-KG matching because systems that rely on matching cell values to KG entities often cannot match GitTables tables.Across the evaluated systems, scores were low; average precision was slightly higher for Schema.org annotations because pattern matching detected a few structural types.Source: §5.3 · p. 15, Figure 6 · p. 15
  • GitTables can support data-search benchmarking: a schema-and-query embedding procedure retrieved a product-order table for “status and sales amount per product.”The authors propose collecting tables and queries and ranking relevant tables to develop the benchmark further.Source: §5.3 · p. 15, Figure 6 · p. 15

6 CONCLUSION

GitTables is a corpus of 1M relational tables extracted from GitHub CSV files, curated for quality, and annotated with semantic column types from DBpedia and Schema.org. Its analyses and applications show complementary value to existing Web-table corpora for detection models, schema completion, and data-management benchmarks.

6 CONCLUSION

  • GitTables contributes a 1M-table relational corpus with semantic annotations, and the paper demonstrates uses in semantic type detection, schema completion, and data-management benchmarks.The authors characterize its value as complementary to existing Web-table corpora.Source: §6 · p. 15

Related papers

Loading 2106.07258v5…