Graph-based Table Representation Learning (GTRL)

We are in the process of updating the functions to align with the revised terminology/concept used in the paper. Consequently, we will soon release our source code and documentation.

News

02/05/2024: Extended results on the public TURL WIKITABLE dataset, associated with the code. Download the dataset and checkpoints "here".

31/12/2023: Initial version of the Graph-based Table Representation Learning (GTRL) project is released. Especially the new benchmarking table dataset of documents about adhesive material. "You can download it here."

Our Dataset

The folder of the "new dataset" about adhesive is structured as follows:

• AdhesiveMaterialsPDF: the collection of PDF documents related to adhesive datasheets under /
• TableDataset: Annotation of each table following the JSON format described below:

{
    "id-table": "0b6979f0-3ffe-431c-83c0-4b3ef6010d31", /*A unique ID assigned to each table, used as the filename. */
    "language": "ENGLISH",  /*The language of the PDF from which the table is extracted.*/
    "table-rect": "(531.7, 294.9, 1011.2, 245.6)", /* The bounding boxes of the table within the PDF, defined by the coordinates (x, y, width, height). */
    "PDF-name": "TDS_3M DP760", /* The name of the PDF file, located in the folder containing the adhesive datasheets.  */
    "Page-number": "2", /*The page number from which this table was extracted. */
    "Cells": [
        {
            "id-entry": 0,  /* The unique id of the cell whitin this table. Note that, this id is idexed and  node id  on the HGT)*/
            "entry-rect": "(544.8, 331.5, 925, 57.5)", /* The bounding boxes for each cell are specified by their (x, y, width, height) coordinates, which are not normalized relative to the table's overall width and height  */
            "table-semantic-label": "KEYWORD", /* The table semantic label is determined based on domain knowledge.  */
            "table-header": "Yes", /* Location of the cell, row-column header or not */
            "content": " Viscosity ", /* The text of the cell */
            "Relationships": {  /* Define relationships using the four cardinal directions - left, right, up, and down */
                "down": [1, 2, 4], /* This cell is identified as a merged cell due to its alignment with multiple cells extending downward. */
                "left": [6], /* The cell id 6 is on its left */
            }
        },
        ...
        // other cell's attributes
        ...
    ]
}

• HeterogeneousGraphTable_HGT-format: the Heterogenous Graph Table (HGT) format saved with gpickle extension using the python Networkx package."

• TABLE_TRAIN-TEST_SPLIT: contains ten random 80%-20% splits for training and testing, respectively, from which the average performance is reported in the study.

• CHECKPOINTS: This folder contains all the model checkpoints, organized by random split IDs.

Replicabilty and reproducibilty

For our dataset (To be completed)

CUDA_VISIBLE_DEVICES=0,1  python3 main.py  --epochs 50 --root_directory "PATH/TO/THE/ADHESIVE_DATASET" -embed_dim 768 --num_hidden_layers 1 --num_attention_heads 12 --save_steps 500 --batch_size 256 --optimizer adam  --smp_target_dataset SMP_CTA_TASK_TSL 

For the TURL dataset:

CUDA_VISIBLE_DEVICES=0,1  python3 main_training_wikidataset_multi_label.py  --epochs 50 --root_directory "PATH/TO/THE/TURL_DATASET" -embed_dim 768 --num_hidden_layers 1 --num_attention_heads 12 --save_steps 500 --batch_size 256 --optimizer adam  --smp_target_dataset SMP_CTA_TASK_TSL 

Model Training

We recommend python3.9 and the necessary packages are listed in the requirements.txt file, while the packages for the entire virtual development environment are detailed in requirements.venv.txt

ASAP

Citing

Our work can be cited using: