MNAV model

1. Prerequisites

First create a conda env which support the required packages use out environment.yml file

conda env create --name mnav --file=environment.yml 

Note: the first line of environment.yml defines the env name.

2. Preprocessing

Getting access to the FewRel and Few-Shot TACRED datasets.

FewRel: https://thunlp.github.io/2/fewrel2_da.html

Few-Shot TACRED: https://github.com/ofersabo/Few_Shot_transformation_and_sampling

Augment the dataset with special tokens before and after the target entities

These commands generate new datasets in which special tokens are surrounding the target entities.

FewRel dataset

python create_dataset_with_marked_entities.py val_wiki.json
python create_dataset_with_marked_entities.py train_wiki.json

Few-Shot TACRED

python create_dataset_with_marked_entities.py data_few_shot/_test_data.json
python create_dataset_with_marked_entities.py data_few_shot/new_downsampled_train_data.json
python create_dataset_with_marked_entities.py data_few_shot/_dev_data.json

3. Generate episodes based on the augmented datasets

FewRel dataset

python FewRel_generate_episodes.py FEWREL_data/train_wiki_markers.json 100000 10 1 50 5 456 FEWREL_data/episodes/train_10w_1s_5q_100K_50_NOTA_rate_seed_456.json 
python FewRel_generate_episodes.py FEWREL_data/train_wiki_markers.json 100000 5 5 50 5 456 FEWREL_data/episodes/train_5w_5s_5q_100K_50_NOTA_rate_seed_456.json
python FewRel_generate_episodes.py FEWREL_data/val_wiki_markers.json 10000 5 1 50 5 456 ./FEWREL_data/episodes/DEV_5w_1s_5q_10K_50_NOTA_rate_seed_456.json
python FewRel_generate_episodes.py FEWREL_data/val_wiki_markers.json 10000 5 5 50 5 456 ./FEWREL_data/episodes/DEV_5w_5s_5q_10K_50_NOTA_rate_seed_456.json

Few-Shot TACRED dataset

To create episodes for Few-shot TACRED, use the script which we published in the Few-Shot TACRED data repository: https://github.com/ofersabo/Few_Shot_transformation_and_sampling

python episodes_sampling_for_fs_TACRED.py --file_name data_few_shot/_dev_data_markers.json --episodes_size 10000 --N 5 --K 1 --number_of_queries 3 --seed 123 --output_file_name TACRED_episodes/dev_5w_1s_3q_10K_seed_123.json
python episodes_sampling_for_fs_TACRED.py --file_name data_few_shot/_dev_data_markers.json --episodes_size 10000 --N 5 --K 5 --number_of_queries 3 --seed 123 --output_file_name TACRED_episodes/dev_5w_5s_3q_10K_seed_123.json 
python episodes_sampling_for_fs_TACRED.py --file_name data_few_shot/new_downsampled_train_data_markers.json --episodes_size 50000 --N 5 --K 1 --number_of_queries 3 --seed 123 --output_file_name TACRED_episodes/train_5w_1s_3q_50K_seed_123.json 
python episodes_sampling_for_fs_TACRED.py --file_name data_few_shot/new_downsampled_train_data_markers.json --episodes_size 50000 --N 5 --K 5 --number_of_queries 1 --seed 123 --output_file_name TACRED_episodes/train_5w_5s_1q_50K_seed_123.json

4. Train the Few-Shot MNAV model

FewRel model

• 1 shot model:

allennlp train experiments/FEWREL_1_shot.jsonnet -s results/fewrel/1_shot/ --include-package my_library

• 5 shot model:

allennlp train experiments/FEWREL_5_shot.jsonnet -s results/fewrel/5_shot/ --include-package my_library

Few-Shot TACRED

• 1 shot model:

allennlp train experiments/Few_Shot_TACRED_1_shot.jsonnet -s results/TACRED/1_shot/MNAV/ --include-package my_library

• 5 shot model:

allennlp train experiments/Few_Shot_TACRED_5_shot.jsonnet -s results/TACRED/5_shot/MNAV/ --include-package my_library