README.md

CORAL

Pytorch implementation of CORAL, with simple annotation.

CORAL: COde RepresentAtion Learning with Weakly-Supervised Transformers for Analyzing Data Analysis

Introduction

We present a new classification task for labeling computational notebook cells as stages in the data analysis process (i.e., data import, wrangling, exploration, modeling, and evaluation). For this task, we propose CORAL, a novel weakly supervised transformer architecture for computing joint representations of data science code from both abstract syntax trees and natural language annotations.

We have shown that CORAL, leveraging only easily-available weak supervision, achieves a 35% increase in accuracy over expert-supplied heuristics. Our model enables us to examine a set of 118,000 Jupyter Notebooks to uncover common data analysis patterns.

This repo is implementation of CORAL. Some code is based on a previous implementation of BERT in pytorch: BERT.

Quickstart

1. Prepare Dataset

Your corpus should be prepared with a code snippet and its parsed AST. See examples in ./examples.
You could generate your own AST with scripts from py-150k.
We use computational notebooks from Data from: Exploration and Explanation in Computational Notebooks

2. Train your own CORAL

./train.sh

3. Test CORAL

./test.sh

Notes

1. Different coverage of weak supervision

Heuristic rules of 20%, 15%, 10% and 5% coverage in ./coral/dataset/seeds.txt. You could replace corresponding codes in ./coral/dataset/dataset.py with them.

2. Annotated results

The annotated test set is available at ./examples/test.txt. It has 1840 cells. You could also use the labels in id2stages.json to create your own dataset.

3. Weak supervision starts before unsupervised topic model

To help the model converge faster, we add unsupervised topic model 1 epoch after weak supervision starts. You could explore this with different values of --hinge_loss_start_point in train.sh.