Observatory

Codebase of paper Observatory: Characterizing Embeddings of Relational Tables (VLDB 2024).

🚀 Update: We have open-sourced the Observatory library for embedding inference of relational tabular data. It is currently released for a beta test and we welcome feedback and/or contributions!

Environment Setup

Assume using Miniconda for Python package management on Linux machines.

Clone this repo in your working directory (the --recursive flag is necessary to pull a dependent repo like TURL as a submodule):
```
git clone <Observatory repo url> --recursive
```
```
cd observatory
```
Create and activate the development environment:
```
conda env create -f environment.yml
```
```
conda activate observatory
```
Note that TaBERT was developed earlier and has many dependencies that are not compatible with other models. We create a separate environment for running TaBERT.
```
conda env create -f tabert.yml
```
```
conda activate tabert
```

Import Observatory and other dependencies as editable packages to the conda environment

conda develop <path to Observatory>

conda develop <path to DODUO>

conda develop <path to TURL>

conda develop <path to TaBERT>

e.g.,

conda develop /home/congtj/observatory

conda develop /home/congtj/observatory/observatory/models/TURL

Models and Data

Model	Checkpoint Identifier	Link
BERT	"bert-base-uncased"	https://huggingface.co/docs/transformers/model_doc/bert
RoBERTa	"roberta-base"	https://huggingface.co/docs/transformers/model_doc/roberta
T5	"t5-base"	https://huggingface.co/docs/transformers/model_doc/t5
TAPAS	"google/tapas-base"	https://huggingface.co/docs/transformers/model_doc/tapas
TaBERT	"TaBERT_Base_(K=3)"	https://github.com/facebookresearch/TaBERT#pre-trained-models
TURL	"checkpoint/pretrained/pytorch_model.bin"	https://github.com/sunlab-osu/TURL#data
DODUO	"wikitable"	https://github.com/megagonlabs/doduo/tree/main#data-preparation

Dataset	Files	Links	Evaluated Properties
WikiTables	data/entity_vocab.txt data/test_tables.jsonl	https://github.com/sunlab-osu/TURL#data	Row Order Insignificance Column Order Insignificance Sample Fidelity
NextiaJD	XS, S, M, L	https://github.com/dtim-upc/NextiaJD/tree/master/experiments#setting	Join Relationship
Dr.Spider	data.tar.gz	https://github.com/awslabs/diagnostic-robustness-text-to-sql	Perturbation Robustness
Spider	Spider Dataset spider_fd_artifact.zip	https://yale-lily.github.io/spider https://drive.google.com/file/d/1br0voV0l3yBMfEy9WM7Vja-cZX6HvvL1/view?usp=sharing	Functional Dependencies
WikiTables_Entities	entity_stability_queries.csv	https://drive.google.com/file/d/1SM_AOpmFbh51IUTQuvI7YETLdjThgpSB/view?usp=sharing	Entity Stability
SOTAB	sotab_data_type_datasets.zip	https://drive.google.com/file/d/1K631KONGDVy2C2ViKcwSyMnWhr_kJdaK/view?usp=share_link	Heterogeneous Context

Evaluation

All property experiments are under observatory/properties. To evaluate a model for a property, change all paths to local ones in the corresponding shell script and run

For models from HuggingFace (specify the model identifier in the script): e.g.,

./observatory/properties/Row_Order_Insignificance/hugging_face.sh

For other models: e.g.,

./observatory/properties/Sample_Fidelity/doduo.sh

Side notes: To evaluate the DODUO model on GPU, please substitute observatory/models/DODUO/doduo/doduo.py with observatory/models/doduo.py.

Use Cases

Besides reproducing the experiments in the paper, we provide a few use cases of Observatory. We are working on cleaning up the codebase and plan to release Observatory as a Python library for easy use. Stay tuned!

1. Embedding Inference

To infer row/column/table embeddings with Hugging Face models for a set of tables, check out get_hugging_face_<row/column/table>_embeddings_batched() in observatory/models/hugging_face_<row/column/table>_embeddings.py.

2. Table Structure Robustness Analysis

To analyze the robustness of a Hugging Face model to table structure changes, e.g., variance of row/column/table embeddings under column order permutations, check out row_embedding_evaluate_col_shuffle.py/evaluate_col_shuffle.py/table_embedding_evaluate_col_shuffle.py in observatory/properties/Column_Order_Insignificance.

3. Join Discovery

For a more extended use case, check out examples/join_discovery/bert/topk_search_nextiaJD.py for join discovery using column embeddings inferred from BERT.

To run the example,

a. Install D3L by running

pip install git+https://github.com/alex-bogatu/d3l

b. Change path in line 9 of examples/join_discovery/bert/topk_search_nextiaJD.py and paths in examples/join_discovery/bert/run_nextiaJD.sh to local ones.

c. Run run_nextiaJD.sh under directory examples/join_discovery/bert/

./run_nextiaJD.sh

Citing This Repository

If you find this repository useful for your work, please cite the following BibTeX:

@article{cong2023observatory,
  author  = {Tianji Cong and
             Madelon Hulsebos and
             Zhenjie Sun and
             Paul Groth and
             H. V. Jagadish},
  title   = {Observatory: Characterizing Embeddings of Relational Tables},
  journal = {Proc. {VLDB} Endow.},
  volume  = {17},
  number  = {4},
  pages   = {849--862},
  year    = {2023},
}

@inproceedings{cong2023observatorylibrary,
  author    = {Cong, Tianji and
               Sun, Zhenjie and
               Groth, Paul and
               Jagadish, H. V. and
               Hulsebos, Madelon},
  title     = {Introducing the Observatory Library for End-to-End Table Embedding Inference},
  booktitle = {The 2nd Table Representation Learning Workshop at NeurIPS 2023},
  publisher = {https://table-representation-learning.github.io},
  year      = {2023}
}

superctj/observatory