Fine Grained Propaganda Detection

We assume that you will be running training/evaluation on your local machine which has multiple GPUs like GTX1080.

pip install -r requirements.txt
For the training, Run one of the following: python train.py --bert --training --batch_size 16 --lr 3e-5 --n_epochs 20 --patience 7 python train.py --joint --training --batch_size 16 --lr 3e-5 --alpha 0.9 --n_epochs 20 --patience 7 python train.py --granu --training --batch_size 16 --lr 3e-5 --alpha 0.9 --n_epochs 20 --patience 7 python train.py --mgn --sig --training --batch_size 16 --lr 3e-5 --alpha 0.9 --n_epochs 20 --patience 7
For the fragment-level evaluation, Run the evaluate.sh ./evaluate.sh ./result/[output.file] bert ./evaluate.sh ./result/[output.file] bert-joint ./evaluate.sh ./result/[output.file] bert-granu ./evaluate.sh ./result/[output.file] mgn
For the span-level evaluation, Run the span-evaluate.sh ./span-evaluate.sh ./result/[output.file] bert ./span-evaluate.sh ./result/[output.file] bert-joint ./span-evaluate.sh ./result/[output.file] bert-granu ./span-evaluate.sh ./result/[output.file] mgn

Tested on:

Python 3.6.8
CUDA 9 Torch 1.0
huggingface/pytorch-pretrained-BERT **0.4 **