SPEAR: Semi suPErvised dAta progRamming

Semi-Supervised Data Programming for Data Efficient Machine Learning

SPEAR is a library for data programming with semi-supervision. The package implements several recent data programming approaches including facility to programmatically label and build training data.

Pipeline

• Design Labeling functions(LFs)
• generate pickle file containing labels by passing raw data to LFs
• Use one of the Label Aggregators(LA) to get final labels

SPEAR provides functionality such as

• development of LFs/rules/heuristics for quick labeling
• compare against several data programming approaches
• compare against semi-supervised data programming approaches
• use subset selection to make best use of the annotation efforts

Labelling Functions (LFs)

• discrete LFs - Users can define LFs that return discrete labels
• continuous LFs - return continuous scores/confidence to the labels assigned

Approaches Implemented

You can read this paper to know about below approaches

• Only-L
• Learning to Reweight
• Posterior Regularization
• Imply Loss
• CAGE
• Joint Learning

Data folder for SMS can be found here. This folder needs to be placed in the same directory as notebooks folder is in, to run the notebooks or examples.

Installation

Method 1

To install latest version of SPEAR package using PyPI:

pip install decile-spear

Method 2

SPEAR requires Python 3.6 or later. First install submodlib. Then install SPEAR:

git clone https://github.com/decile-team/spear.git
cd spear
pip install -r requirements/requirements.txt

Citation

@misc{abhishek2021spear,
      title={SPEAR : Semi-supervised Data Programming in Python}, 
      author={Guttu Sai Abhishek and Harshad Ingole and Parth Laturia and Vineeth Dorna and Ayush Maheshwari and Ganesh Ramakrishnan and Rishabh Iyer},
      year={2021},
      eprint={2108.00373},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Quick Links

• SPEAR tutorials
• SPEAR documentation
• SMS SPAM: CAGE colab, JL colab
• DECILE website
• SubModLib - Summarize massive datasets using submodular optimization
• DISTIL- Deep Diversified Interactive Learning
• CORDS- COResets and Data Subset Selection

Acknowledgment

SPEAR takes inspiration, builds upon, and uses pieces of code from several open source codebases. These include Snorkel, Snuba & Imply Loss. Also, SPEAR uses SUBMODLIB for subset selection, which is provided by DECILE too.

Team

SPEAR is created and maintained by Ayush, Abhishek, Vineeth, Harshad, Parth, Pankaj, Rishabh Iyer, and Ganesh Ramakrishnan. We look forward to have SPEAR more community driven. Please use it and contribute to it for your research, and feel free to use it for your commercial projects. We will add the major contributors here.

Publications

[1] Maheshwari, Ayush, et al. Data Programming using Semi-Supervision and Subset Selection, In Findings of ACL (Long Paper) 2021.

[2] Chatterjee, Oishik, Ganesh Ramakrishnan, and Sunita Sarawagi. Data Programming using Continuous and Quality-Guided Labeling Functions, In AAAI 2020.

[3] Sahay, Atul, et al. Rule augmented unsupervised constituency parsing, In Findings of ACL (Short Paper) 2021.