Skip to content

Latest commit

 

History

History
62 lines (44 loc) · 2.7 KB

README.md

File metadata and controls

62 lines (44 loc) · 2.7 KB
Raw

infoVerse: A Universal Framework for Dataset Characterization with Multidimensional Meta-information

This repository provides datasets, demo and code of the following paper:

infoVerse: A Universal Framework for Dataset Characterization with Multidimensional Meta-information
Jaehyung Kim, Yekyung Kim, Karin de Langis, Jinwoo Shin, Dongyeop Kang
ACL 2023 (main track, long paper)

Installation

The following command installs all necessary packages:

pip install -r requirements.txt

The project was tested using Python 3.7.

Construction of infoVerse

To construct infoVerse, one first needs to 1) train the vanilla classifiers. Then, using the trained classifiers, one can construct infoVerse by extracting the pre-defined meta-information (defined in ./src/scores_src). We release the constructed infoVerse at google drive. Please check out run.sh.

  1. Train the classifiers used for gathering meta-informations
python train.py --train_type 0000_base --save_ckpt --epochs 10 --dataset sst2 --seed 1234 --backbone roberta_large
  1. Construction of infoVerse
python construct_infoverse.py --train_type 0000_base --seed_list "1234 2345 3456" --epochs 10 --dataset sst2 --seed 1234 --backbone roberta_large

In addition, one can visualize the constructed infoVerse and use it to analyize the given dataset using visualize.ipynb. For example, we provide a code to generate an interactive html file, as shown in the below figure. Pre-constructed tSNE and HTML files can be downloaded from the google drive.

Real-world Application #1: Data Pruning

Please see the repository ./data_pruning.

Real-world Application #2: Active Learning

Please see the repository ./active_learning.

Real-world Application #3: Data Annotation

Please see the repository ./data_annotation.

Citation

If you find this work useful for your research, please cite our papers:

@article{kim2023infoverse,
  title={infoVerse: A Universal Framework for Dataset Characterization with Multidimensional Meta-information},
  author={Kim, Jaehyung and Kim, Yekyung and de Langis, Karin and Shin, Jinwoo and Kang, Dongyeop},
  journal={The 61st Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2023}
}