CMDL

Cross-Modal Data Discovery over Structured and Unstructured Data Lakes
Published at Very Large Databses (VLDB) 2023

Set up:

• environment.yml will set up a conda environment

Entry points:

• trainer/pretrain-text.ipynb: Fine tuning a language model on text corpus to learn text embeddings

• trainer/pretrain-tables.ipynb: Fine tuning a language model on table collection to learn tuple embeddings

• trainer/column_text_joint_training.ipynb: training a baseline connecting text to table columns

• compare_gt.py: accuracy measurement of search based baselines and similarity sketches on text->table relation discovery using the ground truth provided

Data Sets & Ground Truths:

All files and directories are inside the inputs directory

• Phamra

  • drugbank-tables: drugbank tables as csv files
  • pubmed-targets: pubmed article abstracts as txt files
  • DrugBank_Synthetic_dataset: synthetic drugbank tables as csv files

• ChEBI

  • ChEBI_tables_dataset: ChEMBL tables as csv files
    Note: chebi-reference.csv.zip & chebi-structures.csv.zip are compressed due to GitHub limits

• ChEMBL

  • ChEMBL_tables_dataset: ChEMBL tables as csv files
    Note: chembl_27-activity_supp.csv.zip , chembl_27-chembl_id_lookup.csv.zip , chembl_27-compound_records.csv.zip , chembl_27-molecule_dictionary.csv.zip are compressed due to GitHub limits

• MLOpen

  • MLOpen Data Source
  • For our experiments we use certain subsets of the data which can be found in the subdirectories:
    • mlopen_t2t_SS_dataset
    • mlopen_t2t_MS_dataset
    • mlopen_t2t_LS_dataset

• UKOpen

  • UKOpen Data Source

The ground truth files for each dataset are present in the inputs directory

Resources:

• Paper manuscripts provided under the folder 'docs'

Prior baselines:

• snorkel labeler.ipynb needs to be run in its separate environment by following instructions at: https://github.com/snorkel-team/snorkel

• build_label_files.py: profiles data, indexes tables, creates labels by probing indexes using each text

• build_features.py: featurizes input data, saves features to disk to be read during training