Domain Adaptation for Disaster Tweet Classification

This repo contains the code for the MVA Deep Learning course project 2020. You can read the report : PR_Castagné.pdf.

It contains code to train a sentence classifier on Disaster Tweet data. In order to make it work, you will need to get the dataset from Alam et al., 2020, available here. All the labels are already in the folder data of this repository.

You can also train RoBERTa adversarially, following Domain Adversarial training for Neural Networks. This is useful to improve robustness of your model. The modified model is in the file model.py, and is mostly inspired by https://github.com/bzantium/bert-DANN.

We offer the option to use Weights and Biases to log experiments. Add the flag --use_wandb to enable it.

Run training of the original classifier

To train with specific sources, use the --train_sources argument with any number of sources you wish. If you want to train with different events, replace that argument by --train_events and specify any event from the dataset you would like.

python main.py --dataset_filename <TRAINING_FILENAME> --model_name_or_path roberta-base --train_sources <TRAIN_SOURCES> --n_epochs 10 --batch_size 32 --lr 1e-5

Run training of the domain adversarial classifier

Training of the adversarial model is similar, except that you need to specify the unlabeled data (target) you want to use, using --target_sources or --target_events.

Be careful : only RoBERTa is available for this training.

python main_adversarial.py --dataset_filename <TRAINING_FILENAME> --train_sources <TRAIN_SOURCES> --target_sources <TARGET_SOURCES> --n_epochs 10 --batch_size 16 --lr 1e-5

Run evaluation of the classifiers on new domains

Add the flag --adversarial to evaluate the adversarial model. In that case, add the target domain on which it was trained on using --target_sources.

python test.py --model_name_or_path roberta-base --train_filename <TRAINING_FILENAME> --test_filename <TRAINING_FILENAME> --train_sources <TRAIN_SOURCES> --test_sources $TARGET_SOURCES