megagonlabs / sato

Code and data for Sato https://arxiv.org/abs/1911.06311.
Apache License 2.0
108 stars 40 forks source link

Sato: Contextual Semantic Type Detection in Tables

This repository includes source code, scripts, and data for training the Sato model. The repo also includes a pretrained model to help replicate the results in our VLDB 2020 paper. Sato is a hybrid machine learning model to automatically detect the semantic types of columns in tables, exploiting the signals from the context as well as the column values. Sato combines a deep learning model trained on a large-scale table corpus with topic modeling and structured prediction.

Above: Sato architecture. Sato's hyrid architecture consists of two basic modules; a topic-aware single-column prediction module and a structured output prediction module. The topic-aware module extends Sherlock's single-column prediction model (a deep neural network) with additional topic subnetworks, incorporating table intent into the model. The structured output prediction module then combines the topic-aware results for all m columns, providing the final semantic type prediction for the columns in the table.

What is Sato useful for?

Myriad data preparation and information retrieval tasks, including data cleaning, integration, discovery and search, rely on the ability to accurately detect data column types. Schema matching for data integration leverages data types to find correspondences between data columns across tables. Similarly, data discovery benefits from detecting types of data columns in order to return semantically relevant results for user queries. Recognizing the semantics of table values helps aggregate information from multiple tabular data sources. Search engines also rely on the detection of semantically relevant column names to extend support to tables. Natural language based query interfaces for structured data can also benefit from semantic type detection.

Demo

We set up a simple online demo where you can upload small tables and get semantic predictions for column types.

screenshot1 screenshot2

Environment setup

We recommend using a python virtual environment:

mkdir virtualenvs
virtualenv --python=python3 virtualenvs/col2type

Fill in and set paths:

export BASEPATH=[path to the repo]
# RAW_DIR can be empty if using extracted feature files.
export RAW_DIR=[path to the raw data]
export SHERLOCKPATH=$BASEPATH/sherlock
export EXTRACTPATH=$BASEPATH/extract
export PYTHONPATH=$PYTHONPATH:$SHERLOCKPATH
export PYTHONPATH=$PYTHONPATH:$BASEPATH
export TYPENAME='type78' 

source ~/virtualenvs/col2type/bin/activate

Install required packages

cd $BASEPATH
pip install -r requirements.txt

To specify GPUID, use CUDA_VISIBLE_DEVICES. CUDA_VISIBLE_DEVICES="" to use CPU.

Replicating results

Results in the paper can be replicated with and pre-trained models features we extracted.

  1. Download data. ./download_data.sh
  2. Run experiments cd $BASEPATH/scripts; ./exp.sh
  3. Generate plots from notebooks/FinalPlotsPaper

Additional

This repo also allows training new Sato models with other hyper-parameters or extract features from additional data.

Download the VIZNET data and set RAW_DIR path to location of VIZNET raw data.

Column feature extraction

cd $BASEPATH/extract
python extract_features.py [corpus_chunk] --f sherlock --num_processes [N]

corpus_chunk: corpus with potential partition post-fix, e.g. webtables0-p1, plotly-p1 N: number of processes used to extract features

Table topic feature extraction

Download nltk data

import nltk
nltk.download('stopwords')
nltk.download('punkt')

[Optional] To train a new LDA model

cd topic_model
python train_LDA.py 

Extract topic features

cd $BASEPATH/extract
python extract_features.py [corpus_chunk] --f topic --LDA [LDA_name] --num_processes [N]

corpus_chunk: corpus with potential partition post-fix, e.g. webtables0-p1, plotly-p1 LDA_name: name of LDA model to extract topic features. Models are located in topic_model/LDA_cache N: number of processes used to extract features

The extracted feature files go to extract/out/features/[TYPENAME] .

Split train/test sets

Split the dataset into training and testing (8/2).

cd $BASEPATH/extract
python split_train_test.py --multi_col_only [m_col] --corpus_list [c_list]

m_col:--multi_col_only is set, filter the result and remove tables with only one column c_list: corpus list

Output is a dictionary with entries ['train','test']. Dictionary values are lists of table_id.

Train Sato

cd $BASEPATH/model
python train_CRF_LC.py -c [config_file]

Check out train_CRF_LC.py for supported configurations.

Original tables

Please see table_data for the original tables used for the experiments.

Citing Sato

Please cite our VLDB 2020 paper

@article{zhang2020sato,
    title={Sato: Contextual Semantic Type Detection in Tables},
    author={Dan Zhang and 
            Yoshihiko Suhara and 
            Jinfeng Li and 
            Madelon Hulsebos and 
            {\c{C}}a{\u{g}}atay Demiralp and 
            Wang-Chiew Tan},
    year = {2020},
    volume = {13},
    number = {12},
    journal = {Proc. VLDB Endow.},
    pages = {1835–1848},
    numpages = {14},
    url = {https://doi.org/10.14778/3407790.3407793}
}

Contact

To get help with problems using Sato or replicating our results, please submit a GitHub issue.