Cyrill Knecht, Joel Neuner-Jehle, Justin Studer, David Zehnder
The paper can be found here.
This is the repository of our proposed sarcasm generation model using self-augmentation. It uses a pretrained discriminator that can distinguish between sarcasm and non-sarcasm to indirectly provide the generator with feedback on the quality of its outputs during training. Outputs classified as sarcasm will be fed back to the generator for further training. The performance of our proposed model is measured for different fractions of mixing self-augmented data with real data and compared to the baseline performance of classic model fine-tuning with and without state-of-the-art data augmentation.
The following figures show an abstract representation of the proposed self-augmentation model and the evaluation procedure:
We provided a reusable ExperimentRunner.ipynb file, runnable e.g. via Google Colaboratory (Pro is required for sufficient resources). This notebook can be used to reproduce all our experiments.
Before running the experiments, additional pretrained models need to be downloaded via provided Google Drive folders. See "/models/classifier/" and "/models/judge/" for respective download links.