Generating Robustness: 6 Ways to Adapt Question Answering to New Domains

CS224N default final project (2022 RobustQA track)

Starter code for robustqa track

• Download datasets from here
• Setup environment with conda env create -f environment.yml
• Train a baseline MTL system with python train.py --do-train --eval-every 2000 --run-name baseline
• Evaluate the system on test set with python train.py --do-eval --sub-file mtl_submission.csv --save-dir save/baseline-01
• Upload the csv file in save/baseline-01 to the test leaderboard. For the validation leaderboard, run python train.py --do-eval --sub-file mtl_submission_val.csv --save-dir save/baseline-01 --eval-dir datasets/oodomain_val

Commands for Data Augmentation

We leverage the nlpaug package that can be found at https:/makcedward/nlpaug for the synonym swapping step, adapted to QA pairs. We leverage the multitask T5 model that can be found at https:/patil-suraj/question_generation, adapted to work with our dataset.

Synthetic data and augmented data that we generated and used for our experiment can be found in the augmented_synthetic_datasets folder. Simply copy all those files into the oodomain_train dataset folder to replicate our results.

We generated these files using the following commands below:

Synthetic Examples
python run_aug.py --synth-file datasets/oodomain_train/duorc

Data Augmentation
python run_aug.py --repeat-aug 1 --aug datasets/oodomain_train/duorc

Combine Synthetic + Augmentation Variants
python run_aug.py --variants synth,aug --combine datasets/oodomain_train/duorc

Commands for Training

• Change training datasets with flags
  • ID: default
  • ID + OOD: --combined
  • ID + OOD augmented: --combinedwAug (need to change datasets)
• Use flag --adv-train for adversarial model
• Use flag --wiki-align for wiki alignment

Example: Train with augmented data and synthetic data, wiki aligned, wasserstein regularization, and lambda annealing python main.py --run-name fancyAdv_wiki_synth_aug1 --do-train --adv-train --combinedwAug --num-epochs 3 --dis-lambda 0.01 --wiki-align --w-reg --anneal

Commands for Finetuning

• Can finetune on OOD, OOD+synth, OOD+aug, or OOD+synth+aug
• Use --variants flag to control what sample is being used for finetuning

Finetuning Example (finetune with both synthetic QA pairs and augmented data) python main.py --do-finetune --finetune-name synth_aug --variants synth,aug --recompute-features --eval-every 50 --num-epochs 3 --eval-dir datasets/oodomain_val --save-dir save/adversarial-baseline

Commands for Evaluation

Eval Example
python main.py --do-eval --finetune-name synth_aug --eval-dir datasets/oodomain_val --save-dir save/adversarial-baseline

Eval Non-finetuned Checkpoint
python main.py --do-eval --eval-dir datasets/oodomain_val --save-dir save/adversarial-baseline

Error Analysis (output predictions with ground truth, question, context, F1, and EM scores) python main.py --do-eval --eval-dir datasets/oodomain_val --save-dir save/adversarial-baseline --error-file error_analysis.csv

Ensemble Evaluation Example python main.py --do-eval --finetune-name aug1,aug1,aug1,synth_aug1 --eval-dir datasets/oodomain_val --save-dir save/adv_wiki_synth_aug1,save/adv_wiki_ood,save/fancyAdv3_multi_synth_aug1,save/adv_wiki_ind