A list of actions and a plan how to solve the challenge.
- explore the data; just look at it
- make a pipeline
- use a dumb baseline (e.g. everything is (not)propaganda)
- evaluate
- replace the dumb model with a finer one
- tune
Source of inspiration: http://karpathy.github.io/2019/04/25/recipe/
Ideas for a better model:
- PyTorch Hub's BERT - https://pytorch.org/hub/huggingface_pytorch-pretrained-bert_bert/
- huggingface's pytorch-transformers with RoBERTa: A Robustly Optimized BERT Pretraining Approach. Paper is here: https://arxiv.org/abs/1907.11692