Datasets:
license: cc-by-4.0
task_categories:
- text-retrieval
language:
- en
size_categories:
- 1K<n<10K
tags:
- table-retrieval
- keyword-search
- data-discovery
- tables
- wikipedia
pretty_name: WikiTables
configs:
- config_name: metadata
data_files:
- split: all
path: metadata.csv
- config_name: queries
data_files:
- split: all
path: queries.csv
- config_name: qrels
data_files:
- split: all
path: qrels.csv
WikiTables
Dataset Summary
WikiTables is a dataset for keyword search over tables (KWS-over-tables): given a short natural-language keyword query (e.g., "2008 beijing olympics"), a retrieval system must rank the tables in a corpus by their relevance to the query. The task is the table analog of text retrieval — instead of returning documents, the system returns structured tables.
The corpus is a downsampled subset of the WikiTables corpus, a large
collection of HTML tables extracted from Wikipedia and originally assembled
for entity linking research. Tables in this corpus do not have schema-style
table names; instead, each table is identified by an opaque table_id and
described by the column names plus surrounding Wikipedia context (page title,
section title, caption).
Relevance judgments are graded on a 0–2 scale (0 = non-relevant,
1 = related, 2 = relevant), inherited unchanged from the original
WikiTables benchmark. This differs from the binary positive-only convention
used by the other datasets in the Polaris suite.
WikiTables is one of six datasets in an evaluation suite for the KWS-over-tables task; each dataset is published as its own Hugging Face repo following a shared schema (with the per-dataset variations noted below).
How to Use
WikiTables is an evaluation-only benchmark — there is no train/test split.
All rows live in a single split named all. (Hugging Face's datasets
library requires every config to declare a split; all is used here in place
of the default train label to avoid implying training data.)
from datasets import load_dataset
# Tables, queries, and graded relevance judgments
metadata = load_dataset("anhaidgroup/wikitables", "metadata", split="all") # 3,361 tables
queries = load_dataset("anhaidgroup/wikitables", "queries", split="all") # 60 queries
qrels = load_dataset("anhaidgroup/wikitables", "qrels", split="all") # 3,120 judgments
For methods that consume tuple-level data, download the per-table tuples archive separately:
from huggingface_hub import hf_hub_download
import zipfile, pandas as pd
zip_path = hf_hub_download("anhaidgroup/wikitables", "tuples.zip", repo_type="dataset")
with zipfile.ZipFile(zip_path) as z:
with z.open("Tuples/table-0001-249.csv") as f:
df = pd.read_csv(f)
Methods that retrieve over table metadata (column names and Wikipedia
context) only need metadata, queries, and qrels. Methods that retrieve
over actual tuple values additionally need tuples.zip.
Dataset Structure
Files
| file | size | purpose |
|---|---|---|
metadata.csv |
3,361 rows | one row per table; identifier, column names, Wikipedia context |
queries.csv |
60 rows | one row per query; identifier and query text |
qrels.csv |
3,120 rows | graded relevance judgments (0/1/2) |
tuples.zip |
3,361 inner CSVs | per-table tuple data, one CSV per table |
Schema
metadata.csv
| column | type | description |
|---|---|---|
table_id |
str | unique identifier, e.g., table-0001-249 |
column_names |
str | JSON-encoded list of column names; parse with json.loads |
table_context |
str | JSON-encoded object with Wikipedia context for the table; commonly contains pgTitle (page title), secondTitle (section title), and caption; parse with json.loads |
There is no table_name column. Wikipedia tables are not named in the source
data — the role usually played by a table name is filled by table_context,
which carries the page and section the table came from.
queries.csv
| column | type | description |
|---|---|---|
query_id |
int | unique identifier, e.g., 1, 2, ..., 60 |
query |
str | natural-language keyword query |
Note that query_id is an integer here, in contrast to the string form
(q1, q2, ...) used by the other Polaris datasets. This matches the
original WikiTables release.
qrels.csv — graded relevance, not positive-only.
| column | type | description |
|---|---|---|
query_id |
int | foreign key to queries.csv |
table_id |
str | identifier from the original WikiTables corpus (see note below) |
relevance_score |
int | 0 = non-relevant, 1 = related, 2 = relevant |
tuples.zip — archive of per-table CSV files.
- Inner layout:
Tuples/<table_id>.csv(one file per table). - Each inner CSV: header on row 0 (matching the
column_namesentry inmetadata.csvfor thattable_id), tuples on subsequent rows, RFC 4180 quoting. - Headers may contain duplicate or empty strings (Wikipedia tables sometimes
have repeated or unnamed columns). Use a CSV parser that returns headers
verbatim;
pandas.read_csvwill auto-suffix duplicates (Foo→Foo.1) and rename blanks (""→Unnamed: 0).
Statistics
- 3,361 tables, after corpus downsampling (see below).
- 60 keyword queries (the full original WikiTables query set).
- 3,120 graded judgments: 2,269 non-relevant (
0), 474 related (1), 377 relevant (2). - Per-query judgment counts range from 39 to 62 (median 52).
Dataset Creation
Source Data
The corpus is a downsampled subset of the WikiTables corpus, a collection of roughly 1.6M HTML tables extracted from English Wikipedia, originally assembled for entity linking research. The 60 keyword queries and the graded (0–2) relevance judgments used here are from the ad hoc table retrieval benchmark introduced by Zhang and Balog (2018) over that corpus. Tables carry column names and surrounding page context (page title, section title, caption) but no schema-style table names.
Downsampling
Generating LLM enrichments for the full 1.6M-table corpus is prohibitively expensive. To make LLM-based KWS evaluation tractable while preserving task difficulty, the corpus is downsampled as follows:
For each query, retrieve the top-50 tables via BM25 and take the union with all tables judged relevant for that query. This focuses the corpus on hard cases (BM25-retrievable distractors) without removing relevant material.
This procedure yields 3,361 retained tables, listed in metadata.csv,
with their full tuple content in tuples.zip.
Annotations
Relevance judgments are inherited unchanged from the Zhang and Balog (2018)
ad hoc table retrieval benchmark and were produced by manual inspection. The
0–2 scale reflects three judgment levels (non-relevant / related / relevant)
and is preserved as-is in qrels.csv rather than being collapsed to binary,
so users can replicate prior published numbers and compute graded-relevance
metrics (e.g., nDCG) directly.
Considerations for Using the Data
Why qrels.csv references some tables not in metadata.csv
The original WikiTables qrels file contains 3,120 (query, table) judgments
covering tables in the full 1.6M-table source corpus. After downsampling,
3,361 of those tables remain in metadata.csv. To preserve the full source
qrels file as published, qrels.csv retains all 3,120 original rows,
including rows whose table_id is not present in metadata.csv.
In practice:
- All
relevance_score = 1andrelevance_score = 2rows reference tables that are inmetadata.csv. Every relevant table was kept by the downsampling procedure (it is in the union of "BM25 top-50 ∪ all relevant tables"). - Rows whose
table_idis not inmetadata.csvall haverelevance_score = 0. These judgments concern tables that were dropped from the corpus during downsampling and cannot be retrieved by any system evaluated on this release.
This is a deliberate choice to preserve the original WikiTables qrels file
verbatim. It does mean that qrels.csv does not satisfy a strict foreign-key
constraint with metadata.csv. Code that joins the two files should either
use a left join (and treat the gap as expected) or filter to in-corpus
judgments before evaluation:
import pandas as pd
md = pd.read_csv("metadata.csv")
ql = pd.read_csv("qrels.csv")
ql_in_corpus = ql[ql.table_id.isin(md.table_id)] # 1,141 rows: 290 zeros + 474 ones + 377 twos
For most retrieval evaluations the in-corpus subset is the right one to use, since a system cannot retrieve a table that is not in the corpus.
Why queries.csv retains all 60 queries
Three queries (12 running shoes, 52 erp systems price, 53 cats
life span) have no relevance_score >= 1 row anywhere in the WikiTables
qrels file — i.e., they had no relevant tables even in the full 1.6M-table
source corpus. We retain these queries in queries.csv to preserve the
original WikiTables query set. Users who want to evaluate only on queries
with at least one relevant table in the downsampled corpus should filter:
qr = pd.read_csv("queries.csv")
ql = pd.read_csv("qrels.csv")
md = pd.read_csv("metadata.csv")
qids_with_relevant = ql[(ql.relevance_score >= 1) & (ql.table_id.isin(md.table_id))].query_id.unique()
qr_eval = qr[qr.query_id.isin(qids_with_relevant)] # 57 queries
Tuple downsampling
This release additionally caps each table's tuples at 500,000 rows
(df.head(500_000) against the source); tables with fewer rows are
unaffected. WikiTables tables are typically small (extracted from
single-page HTML), so this cap binds rarely if at all.
metadata.csv and qrels.csv were produced against the un-capped corpus.
Methods that rely only on metadata are unaffected; methods that inspect
tuple values may produce a slightly deflated score relative to the same
method evaluated on the un-capped corpus.
License
The dataset is released under CC-BY-4.0. You are free to share and adapt the material with attribution.
The source WikiTables corpus is derived from English Wikipedia content; attribution should be given to Wikipedia and to the original WikiTables release.
Citation
A citation for this dataset release will be added once the associated paper is published.
The queries and graded relevance judgments are from:
Zhang, S., and Balog, K. (2018). Ad hoc table retrieval using semantic similarity. In Proceedings of the 2018 World Wide Web Conference (WWW '18), pp. 1553–1562.
@inproceedings{zhang2018adhoc,
author = {Zhang, Shuo and Balog, Krisztian},
title = {Ad Hoc Table Retrieval using Semantic Similarity},
booktitle = {Proceedings of the 2018 World Wide Web Conference},
series = {WWW '18},
year = {2018},
pages = {1553--1562}
}
Authors
Minh Phan, Ting Cai, AnHai Doan.