To run a generation experiment (either conceptnet or atomic), follow these instructions:

First Steps

First clone, the repo:

git clone https://github.com/atcbosselut/comet-commonsense.git

Then run the setup scripts to acquire the pretrained model files from OpenAI, as well as the ATOMIC and ConceptNet datasets

bash scripts/setup/get_atomic_data.sh
bash scripts/setup/get_conceptnet_data.sh
bash scripts/setup/get_model_files.sh

Then install dependencies (assuming you already have Python 3.6 and Pytorch >= 1.0:

pip install tensorflow
pip install ftfy==5.1
conda install -c conda-forge spacy
python -m spacy download en
pip install tensorboardX
pip install tqdm
pip install pandas
pip install ipython

Making the Data Loaders

Run the following scripts to pre-initialize a data loader for ATOMIC or ConceptNet:

python scripts/data/make_atomic_data_loader.py
python scripts/data/make_conceptnet_data_loader.py

For the ATOMIC KG, if you'd like to make a data loader for only a subset of the relation types, comment out any relations in lines 17-25.

For ConceptNet if you'd like to map the relations to natural language analogues, set opt.data.rel = "language" in line 26. If you want to initialize unpretrained relation tokens, set opt.data.rel = "relation"

Setting the ATOMIC configuration files

Open config/atomic/changes.json and set which categories you want to train, as well as any other details you find important. Check src/data/config.py for a description of different options. Variables you may want to change: batch_size, learning_rate, categories. See config/default.json and config/atomic/default.json for default settings of some of these variables.

Setting the ConceptNet configuration files

Open config/conceptnet/changes.json and set any changes to the degault configuration that you may want to vary in this experiment. Check src/data/config.py for a description of different options. Variables you may want to change: batch_size, learning_rate, etc. See config/default.json and config/conceptnet/default.json for default settings of some of these variables.

Running the ATOMIC experiment

Training

For whichever experiment # you set in config/atomic/changes.json (e.g., 0, 1, 2, etc.), run:

python src/main.py --experiment_type atomic --experiment_num #

Evaluation

Once you've trained a model, run the evaluation script:

python scripts/evaluate/evaluate_atomic_generation_model.py --split $DATASET_SPLIT --model_name /path/to/model/file

Generation

Once you've trained a model, run the generation script for the type of decoding you'd like to do:

python scripts/generate/generate_atomic_beam_search.py --beam 10 --split $DATASET_SPLIT --model_name /path/to/model/file
python scripts/generate/generate_atomic_greedy.py --split $DATASET_SPLIT --model_name /path/to/model/file
python scripts/generate/generate_atomic_topk.py --k 10 --split $DATASET_SPLIT --model_name /path/to/model/file

Running the ConceptNet experiment

Training

For whichever experiment # you set in config/conceptnet/changes.json (e.g., 0, 1, 2, etc.), run:

python src/main.py --experiment_type conceptnet --experiment_num #

Development and Test set tuples are automatically evaluated and generated with greedy decoding during training

Generation

If you want to generate with a larger beam size, run the generation script

python scripts/generate/generate_conceptnet_beam_search.py --beam 10 --split $DATASET_SPLIT --model_name /path/to/model/file

Classifying Generated Tupes

To run the classifier from Li et al., 2016 on your generated tuples to evaluate correctness, first download the pretrained model from:

wget https://ttic.uchicago.edu/~kgimpel/comsense_resources/ckbc-demo.tar.gz
tar -xvzf ckbc-demo.tar.gz

then run the following script on the the generations file, which should be in .pickle format:

bash scripts/classify/classify.sh /path/to/generations_file/without/pickle/extension

If you use this classification script, you'll also need Python 2.7 installed.

Playing Around in Interactive Mode

First, download the pretrained models from the following link:

https://drive.google.com/open?id=1FccEsYPUHnjzmX-Y5vjCBeyRt1pLo8FB

Then untar the file:

tar -xvzf pretrained_models.tar.gz

Then run the following script to interactively generate arbitrary ATOMIC event effects:

python scripts/interactive/atomic_single_example.py --model_file pretrained_models/atomic_pretrained_model.pickle

Or run the following script to interactively generate arbitrary ConceptNet tuples:

python scripts/interactive/conceptnet_single_example.py --model_file pretrained_models/conceptnet_pretrained_model.pickle

Bug Fixes

Beam Search

In BeamSampler in sampler.py, there was a bug that made the scoring function for each beam candidate slightly different from normalized loglikelihood. Only sequences decoded with beam search are affected by this. It's been fixed in the repository, and seems to have little discernible impact on the quality of the generated sequences. If you'd like to replicate the exact paper results, however, you'll need to use the buggy beam search from before, by setting paper_results = True in Line 251 of sampler.py

References

Please cite this repository using the following reference:

@inproceedings{Bosselut2019COMETCT,
  title={COMET: Commonsense Transformers for Automatic Knowledge Graph Construction},
  author={Antoine Bosselut and Hannah Rashkin and Maarten Sap and Chaitanya Malaviya and Asli Çelikyilmaz and Yejin Choi},
  booktitle={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2019}
}