solution write-up is at: https://www.kaggle.com/c/gendered-pronoun-resolution/discussion/90334

paper is at: https://arxiv.org/abs/1905.01780

File descriptions

• Step1_preprocessing.ipynb given [dev, test, val, stage2] input tsv files, generate augmented tsv; extract bert features jsons, distance features, and linguistic features, saved to Drive

• Step2_end2end_model.ipynb from [dev, test, val, stage2] features saved in step 1, train "end2end" model and save weights (250 for sub_A and 50 for sub_B) -- only used for training

• Step3_pure_bert_model.ipynb from [dev, test, val, stage2] features saved in step 1, train "pure bert" model and save weights (50 for sub_A and 50 for sub_B) -- only used for training

• Step4_inference.ipynb using saved features in step1 and saved weigths in step2 and 3, do inference

• gap-development-corrected-74.tsv dev input file with 74 wrong labeled fixed -- only used for training

• gap-test-val-85.tsv test and val input file (combined) with 85 wrong labeled fixed -- only used for training

Training insturctions

1. Run Step1_preprocessing.ipynb to generate augmented tsv files and extract features for dev, test and val
2. Run Step2_end2end_model.ipynb to train the "end2end" model and save weights
3. Run Step3_pure_bert_model.ipynb to train the "pure bert" model and save weights

Both step2 and 3 need to be ran 4 times, with all_train = True and False, and CASED = False and True respectively. Step2 and 3 will generate 5.11G weights files in total.

Inference instructions

1. Run Step1_preprocessing.ipynb to extract features for stage2 data
2. Run Step4_inference.ipynb twice to do inference for both subissions: all_train=True for sub_A, all_train=False for sub_B