(Comet-) ATOMIC 2020
(Comet-) ATOMIC 2020: On Symbolic and Neural Commonsense Knowledge Graphs
Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, Yejin Choi
"(Comet-) Atomic 2020: On Symbolic and Neural Commonsense Knowledge Graphs."
Appearing at AAAI Conference on Artificial Intelligence 2021
Data: ATOMIC 2020
Model: COMET-ATOMIC 2020
Trained COMET model can be downloaded here.
We include code used in expirements in COMET-ATOMIC2020 for reproducibility, ease of use. Our models are based off the HuggingFace Transformers codebase, with minor adjustments to adapt the model for our data. Details can be found in the AAAI paper.
pip install -r requirements.txt to install requirements for your Python instance. We recommend Conda to manage Python installs. Our codebases is on Python 3.
It's recommended that you test that your enviroment is set up correctly before running modeling code. You can do this via
python models/comet_atomic2020_gpt2/comet_gpt2.py --test_install
The code for modeling is located in
mosaic/datasets/KGDataset is used to convert the ATOMIC2020 CSV into an HuggingFace
beaker_exp: Contains files needed to run expirements using
Beaker (https://beaker.org/) instead of on your local machine.
human_eval: Contains HTML files for human evaluation on Amazon MTurk, as described in the AAAI paper.
models: Contains additional modeling files to reproduce the GPT2 and BART expirements.
models/comet_atomic2020_bart contains a README and code to run COMET-BART2020.
scripts: Contains additional scripts (e.g. utils.py) used during expirements in the COMET-ATOMIC2020 paper.
split: Contains code used to make the test, train, and dev splits of ATOMIC2020 with Stratified Random Sampling.
system_eval: Contains code for automatic evaluation of generated entities.
We welcome contributions to the codebase of COMET-2020. We encourage pull requests instead of issues; and suggest filing a GitHub issue with questions / suggestions.