Propaganda Analysis in News Articles
Online media articles often have source induced biases that sway user opinions and perspectives. There is no system in common knowledge with explainable decisions that identifies and removes these, often subjective, biases and can be used across sources. In this work we have made progress towards making an end to end framework for Fine Grained detection of propaganda in News Articles and then Rewriting them with a Neutral Point of view.
Getting Started
To run the code for training with BERT as backbone simply clone the repository and run
python3 train.py --training --bert
Additional Parameters Used are:
| Argument | Default Value |
|---|---|
| Batch Size | 16 |
| Learning Rate | 3*10-5 |
| Group Classes | True |
| Device | Cuda |
Prerequisites
Pytorch <=1.40wandb
Different Experiments and Architectures
- BERT Model 18 Class output(Both sentence and Token Level Classification)
- BERT + CRF
- BERT Grouped into 3 Classes of Bias
- BERT + Auxiliary objective of Valence Arousal and Dominance prediction
Running the tests
Set the Training Flag to False and change the input path to the Dev/Test Dataset
Authors
- Chinmay Singh - chinmay-singh
- Ayush Kaushal - ayushk4
Acknowledgments
- Vitobha Munigala, IBM Research
- Nishtha Madan, IBM Research