AddressTable

AddressTable: A Large-Scale Real-World Address Dataset for Data Cleaning Benchmarking This dataset accompanies the paper: Clean Me If You Can: A Large Collection of Real-World Addresses for Data Cleaning BenchmarkingFatemeh Ahmadi, Tobias Bernhard, Mohamed Abdelmaksoud, Luca Zecchini, Tilmann Rabl, Ziawasch Abedjan GitHub Repo, arxiv paper Overview AddressTable is a large-scale, real-world address dataset with validated ground truth for benchmarking cell-level error detection and correction in tabular data. It contains postal addresses of public entities (businesses, hotels, libraries, hospitals, etc.) extracted from the Web Data Commons (WDC) corpus. Each address is paired with a ground-truth record obtained by matching it to OpenStreetMap via Nominatim, then validated through a manual audit and corrected using authoritative national address registers. The dataset captures naturally occurring data quality issues across 128 countries: missing values, abbreviations, language variations, encoding artifacts, typos, factually wrong values, and more. Dataset Contents The archive `AddressTable.zip` contains the datasets below. Every dataset is provided as a dirty version and a ground-truth version with the same schema and a one-to-one row correspondence. Full Versions (Addresses) Dataset # Records Description full-named 4,335,421 Addresses with entity names (e.g., hotel or library name). Recommended for general-purpose evaluation. full-unnamed 9,312,725 Addresses without entity names Country Slices (from `full-named`) Slice # Records us-slice 2,267,134 de-slice 374,503 ru-slice 194,229 POI Slices 22 point-of-interest slices reflecting Schema.org entity types (e.g., `library`, `shopping-center`, `hotel`, `hospital`). The error-type distribution of every slice is reported in the repository, so you can pick slices that emphasize specific error types. See the paper and repository for the full list. ### Benchmark Samples Samples used in the experiments reported in the paper: - Effectiveness: 10,000-row samples of `full-named`, `de-slice`, and `ru-slice` at error rates of 5%, 10%, 20%, and the original rate; plus the full `library` (1,656 rows) and `shopping-center` (802 rows) samples at their original rates.- Scalability: random `full-named` samples of 10k, 100k, 500k, and 1M records, plus 10k and 100k samples containing only distinct entities. Customization Scripts The repository provides scripts that let you adapt the benchmark to your definition of an error: - exclude selected columns (e.g., `country_code`, which is 75.1% missing);- exclude selected error types (e.g., purely representational errors such as abbreviations). Cells with excluded error types are restored to their ground-truth value in the dirty table. Schema Attribute Description _id Unique record identifier name Entity name (full-named version only) house_number House/building number road Street name postcode Postal code locality City, town, or village region State, province, or administrative region country Country name country_code ISO 3166-1 alpha-2 country code graph_iri Source page URL (provenance) Ground Truth The ground truth was created in four steps: Matching. Each dirty record was queried against a self-hosted Nominatim instance (OpenStreetMap). Records without a match were discarded, and records with incomplete ground truth were removed. For full-named, the entity name was additionally required to match exactly (token set ratio of 100%) after normalization. Manual audit. Five annotators checked 11,220 records with at least one erroneous cell (stratified by country and error type) against official sources. They verified 96.5% of the cells and corrected 1.8%. The remaining 1.7% were disqualified because the matched entity could not be verified, and their records were removed. Because the audit targets conflicting records, these figures are conservative estimates for the full dataset. Systematic corrections. Country-specific OSM errors found in the audit were corrected across the full dataset using national sources, e.g., the NAD and NAR address registers for US and Canadian postcodes, and postcode-based lookups for coarse regions in France and coarse localities in Great Britain and Australia. This corrected 637,529 cells (full-named) and 1,250,122 cells (full-unnamed), and removed 2,688 and 5,161 records, respectively. Error Statistics In `full-named`, **45.8%** of cells contain errors (54.2% are clean); in `full-unnamed`, 50.1%. Values that differ only in letter case, whitespace, or Unicode normal form are treated as clean. Errors are classified by a rule-based **multi-label** classifier. A cell receives every category that applies, e.g., `Mainn Ave` ↔ `Main Avenue` is labeled both *Abbreviation* and *Typo*. There is one *Clean* category and eight error categories: Category Description Example Missing Empty value `""` ↔ `"Berlin"` Special Charatcters Punctuation differences `"U.S.A."` ↔ `"USA"` Formatting Token reordering, number notation, or value formats `"Howson Branch, Austin Public Library"` ↔ `"Austin Public Library – Howson Branch"` Language Translation, exonym, or latinization `"Austria"` ↔ `"Österreich"` Abbreviation Shortened forms `"Ave"` ↔ `"Avenue"`, `"Bachmannstr."` ↔ `"Bachmannstraße"` Token Subset Partial names `"New York"` ↔ `"New York City"` Typo Small edit-distance differences `"Theatre"` ↔ `"Theater"` Semantic Valid but factually wrong value, regardless of formatting `"Lich-Steinstraß"` ↔ `"Jülich"`