To reproduce the results from the paper, you can direclty use the pretrained models or extracted features feature we provided in the main download script. Here we also provide the raw and processed table we used in our experiements. viznet_tables is a subset of the VizNet corpus, and sato_tables contains the same set of tables but filtered to only include columns with valid semantic types. Please see below for dataset details.
Run the script and you will see sata_tables and viznet_tables.
$ bash download.sh
The directory contains two directories all and multionly, which correspond to All tables D and Multi-column tables D_multi in the paper. The tables are split into 5 disjoint subset for cross-validation. Each table is stored in a CSV format, so you can load the files using a standard way (e.g., Pandas read_csv().)
sato_tables/
└── all
├── K0
├── K1
├── K2
├── K3
└── K4
└── multionly
├── K0
├── K1
├── K2
├── K3
└── K4
For each CV, the same block id as the cv index was used as the test data. That is,
CV0: Train K1, 2, 3, 4, Test K0
CV1: Train K0, 2, 3, 4, Test K1
CV2: Train K0, 1, 3, 4, Test K2
CV3: Train K0, 1, 2, 4, Test K3
CV4: Train K0, 1, 2, 3, Test K4
In case you're interestd in using the original tables that are extracted from the VizNet corpus before filtering and semantic type assignment, please refer to viznet_tables.
viznet_tables stores the original table files in a CSV format. K{0-4} contains all the original tables and K{0-4}_multi-col contains a subset of the tables that have more than one columns (i.e., multi-column tables in the experiments.)
viznet_tables/
├── webtables1
│ ├── K0
│ ├── K0_multi-col
│ ├── K1
│ ├── K1_multi-col
│ ├── K2
│ ├── K2_multi-col
│ ├── K3
│ ├── K3_multi-col
│ ├── K4
│ └── K4_multi-col
└── webtables2
├── K0
├── K0_multi-col
├── K1
├── K1_multi-col
├── K2
├── K2_multi-col
├── K3
├── K3_multi-col
├── K4
└── K4_multi-col