/MoCA

Reproducing Momentum Calibration for Text Generation Paper

Primary LanguageJupyter Notebook

Reproducing Momentum Calibration for Text Generation Paper

Part of onboarding task for Forward Data Lab Summer 2023, I reproduced the current state-of-the-art abstractive text summerication method: Momentum Calibration for Text Generation (MoCA) using huggingface and pytorch. See Final Report for this task.

Results:

Rouge 1 Rouge 2 Rouge L
Vanilla Finetuned Baseline in the MoCa Paper 44.22 21.22 41.01
MoCa Paper 48.88 24.94 45.76
Vanilla Finetuned Baseline in the BART paper 40.95 20.81 40.04
My Implementation 46.04 22.61 42.93

Running the code

  • Set/Change the environment variable for selecting GPUs to use.
  • Run python MoCa.py to apply MoCA on model.
  • Run evaluation.ipynb to evaluate model on ROUGE score.