README.md

---
license: cc-by-4.0
task_categories:
  - text-retrieval
language:
  - en
size_categories:
  - n<1K
tags:
  - table-retrieval
  - keyword-search
  - data-discovery
  - tables
pretty_name: AW
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
---

# AW

## Dataset Summary

AW is a dataset for **keyword search over tables (KWS-over-tables)**: given a
short natural-language keyword query (e.g., *"employee phone number"*), 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 — and is motivated by
enterprise data discovery, where analysts who know roughly what they need
must locate the right table in a database with many cryptically named ones.

The corpus is a modified version of the AdventureWorks Microsoft sample
database. Column names have been manually abbreviated (e.g., `BusEntId`,
`DTPh`, `STRGUID`) to simulate the low-information schemas common in real
enterprise data.

Relevance judgments are **binary (0 or 1)** and were produced by manual
inspection of every (query, table) pair in the full 96 × 15 = 1,440 Cartesian
product. The 60 pairs marked relevant (`1`) are listed in `qrels.csv`; pairs
not listed are non-relevant.

AW is one of six datasets in an evaluation suite for this task; each dataset
is published as its own Hugging Face repo following the same schema.

## How to Use

AW 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.)

```python
from datasets import load_dataset

# Tables, queries, and binary relevance judgments
metadata = load_dataset("anhaidgroup/aw", "metadata", split="all")  # 96 tables
queries  = load_dataset("anhaidgroup/aw", "queries",  split="all")  # 15 queries
qrels    = load_dataset("anhaidgroup/aw", "qrels",    split="all")  # 60 relevant pairs
```

For methods that consume tuple-level data, download the per-table tuples
archive separately:

```python
from huggingface_hub import hf_hub_download
import zipfile, pandas as pd

zip_path = hf_hub_download("anhaidgroup/aw", "tuples.zip", repo_type="dataset")
with zipfile.ZipFile(zip_path) as z:
    with z.open("Tuples/Sales-17.csv") as f:
        df = pd.read_csv(f)
```

Methods that retrieve over table metadata (table name and column names) 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` | 96 rows | one row per table; identifier, name, column names |
| `queries.csv` | 15 rows | one row per query; identifier and query text |
| `qrels.csv` | 60 rows | binary relevance judgments (positive pairs only) |
| `tuples.zip` | 96 inner CSVs | per-table tuple data, one CSV per table |

### Schema

`metadata.csv`

| column | type | description |
|---|---|---|
| `table_id` | str | unique identifier, e.g., `Person-0`, `Sales-17` |
| `table_name` | str | name from the source schema |
| `column_names` | str | JSON-encoded list of column names; parse with `json.loads` |

`queries.csv`

| column | type | description |
|---|---|---|
| `query_id` | str | unique identifier, e.g., `q1`, `q2` |
| `query` | str | natural-language keyword query |

`qrels.csv` — positive-only convention: only relevant (query, table) pairs
are listed. Pairs not listed are treated as non-relevant.

| column | type | description |
|---|---|---|
| `query_id` | str | foreign key to `queries.csv` |
| `table_id` | str | foreign key to `metadata.csv` |
| `relevance_score` | int | always `1` (binary judgments) |

`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_names` entry in
  `metadata.csv` for that `table_id`), tuples on subsequent rows, RFC 4180
  quoting.
- Some tables (e.g., `Production-2`) contain multi-line quoted text fields.
  Use a CSV parser that respects quoting (`pandas.read_csv` works); avoid
  line-based tools.

### Statistics

96 tables span six namespaces from the source schema: `Person` (13),
`Sales` (19), `dbo` (28), `Production` (25), `Purchasing` (5),
`HumanResources` (6). Per-query relevant-table counts range from 1 to 8
(median 4).

## Dataset Creation

### Source Data

The corpus is built from a modified version of AdventureWorks, Microsoft's
sample relational database for SQL Server demos. The schema modification is
the manual abbreviation of column names — tuple values are unchanged from
the original. This modified schema was originally introduced in the Columbo
system to study retrieval and discovery on schemas with low-information
names.

### Annotations

All 1,440 candidate (query, table) pairs (96 tables × 15 queries) were
manually inspected. A pair was marked relevant (`1`) if the table contains
data that would satisfy the query, judged from the table name, column names,
and a sample of tuple values. Pairs not marked are non-relevant. The
positive-only convention (60 entries listed in `qrels.csv`, the rest treated
as non-relevant by absence) follows TREC qrels practice.

## Considerations for Using the Data

### Tuple downsampling

This release caps each table's tuples at 500,000 rows, applied as
`df.head(500_000)` against the source — the first 500k rows of each table
preserving the original ordering. Tables with fewer than 500k rows are
unaffected.

`metadata.csv` and `qrels.csv` were produced against the un-capped corpus.
Methods that rely only on table metadata (table name and column names) are
unaffected by the cap. Methods that inspect actual tuple values may produce
a deflated score relative to the same method evaluated on the un-capped
corpus, since some relevant cells in tail rows of large tables may not
appear in `tuples.zip`. A future release will include the un-capped tuples.

## License

The dataset is released under
[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). You are free to
share and adapt the material with attribution.

The source AdventureWorks database is a Microsoft sample.

## Citation

A citation will be added once the associated paper is published.

<!-- TODO: replace with BibTeX entry on publication -->

## Authors

Minh Phan, Ting Cai, AnHai Doan.