Demo

Please download pretrained weights beforehand.

Requirements

python 3.6

To install all requirements, run pip install -r requirements.txt

Data Format

The code assumes the input text is in a json-formatted file with the following keys:

'id': unique id for each narrative

'full_context': the narrative split into a list of sentences

Generation

Para-M: python demo.py --model_dir /path/to/modeldir/ --model_type 'nomem' --decoding beam --beam 10 --source example.jsonl

Para-M (mem): python demo.py --model_dir /path/to/modeldir/ --model_type 'mem' --decoding beam --beam 10 --source example.jsonl

Citing

@article{Gabriel2021ParagraphLevelCT,

  title={Paragraph-level Commonsense Transformers with Recurrent Memory},
  
  author={Saadia Gabriel and Chandra Bhagavatula and Vered Shwartz and Ronan Le Bras and M. Forbes and Yejin Choi},
  
  journal={AAAI},
  
  year={2021}

Note

For now, code must be run on a single GPU.