jboru / hykas-cskg

Code for the paper "Knowledge-driven Data Construction for Zero-shot Evaluation in Commonsense Question Answering" (AAAI 2021)

Knowledge-driven Data Construction for Zero-shot Evaluation in Commonsense Question Answering

This repository contains the code for the paper "Knowledge-driven Data Construction for Zero-shot Evaluation in Commonsense Question Answering" (AAAI 2021). See full paper here

Note that our evaluation code is adpated from self-talk repo

Enviroments

This code has been tested on Python 3.7.6, Pytorch 1.5.1 and Transformers 3.0.2, you can install the required packages by

pip install -r requirements.txt

Data generation

Our synthetic QA sets can be downloaded from here, uncompress it and place it in the HyKAS-CSKG root directory.

If you would like to generate data from scratch, first cd to the src/Data_generation directory.

For the ATOMIC synthetic sets, download the ATOMIC from official website and uncompress. Then run

python generate_from_ATOMIC.py --train_KG atomic/v4_atomic_trn.csv --dev_KG atomic/v4_atomic_dev.csv --strategy random --out_dir ../../data/ATOMIC  

For CWWV, download the cskg_connected.tsv from here and cache.pkl from here, then run:

python generate_from_CWWV.py --cskg_file cskg_connected.tsv --lex_cache cache.pkl --output_dir ../../data/CWWV --strategy random
python filter_CWWV.py --input_file ../../data/CWWV/random.jsonl 

Pretraining on Synthetic QA sets

We provide following pretrained models

LM	KG	Download
RoBERTa-Large	ATOMIC	Download
RoBERTa-Large	CWWV	Download
RoBERTa-Large	CSKG	Download
GPT2-Large	ATOMIC	Download
GPT2-Large	CWWV	Download
GPT2-Large	CSKG	Download

If you would like to train models from scratch, you can use the following commands under src/Training

For RoBERTa

CUDA_VISIBLE_DEVICES=0 python run_pretrain.py --model_type roberta-mlm --model_name_or_path roberta-large --task_name cskg --output_dir ../../out_dir --max_sequence_per_time 200 \
--train_file ../../data/ATOMIC/train_random.jsonl --second_train_file ../../data/CWWV/train_random.jsonl --dev_file ../../data/ATOMIC/dev_random.jsonl --second_dev_file \
../../data/CWWV/dev_random.jsonl --max_seq_length 128 --max_words_to_mask 6 --do_train --do_eval --per_gpu_train_batch_size 2 --gradient_accumulation_steps 16 \
--learning_rate 1e-5 --num_train_epochs 1 --warmup_proportion 0.05 --evaluate_during_training --per_gpu_eval_batch_size 8  --save_steps 6500 --margin 1.0

For GPT2

CUDA_VISIBLE_DEVICES=0 python run_pretrain_gpt2.py --model_type gpt2 --model_name_or_path gpt2-large --task_name cskg --output_dir ../../out_dir --train_file ../../data/ATOMIC/ \
train_random.jsonl --second_train_file ../../data/CWWV/train_random.jsonl --dev_file ../../data/ATOMIC/dev_random.jsonl --second_dev_file ../../data/CWWV/dev_random.jsonl \
--max_seq_length 128 --do_train --do_eval --per_gpu_train_batch_size 2 --gradient_accumulation_steps 16 --learning_rate 1e-5 --num_train_epochs 1 --warmup_proportion 0.05 \
--evaluate_during_training --per_gpu_eval_batch_size 8  --save_steps 6500 --margin 1.0

Evaluation

For LM baselines, cd to src/Evaluation directory

python evaluate_RoBERTa.py --lm roberta-large --dataset_file DATA_FILE --out_dir ../../results --device 1 --reader TASK_NAME
python evaluate_GPT2.py --lm gpt2-large --dataset_file DATA_FILE --out_dir ../../results --device 1  --reader TASK_NAME

For pretrained models, simply point the --lm flag to your model directory, for example

python evaluate_RoBERTa.py --lm ../../models/roberta_cskg --dataset_file ../../tasks/commonsenseqa_dev.jsonl --out_dir ../../results --device 1 --reader commonsenseqa
python evaluate_GPT2.py --lm ../../models/gpt2_cskg --dataset_file ../../tasks/commonsenseqa_dev.jsonl --out_dir ../../results --device 1  --reader commonsenseqa

MLM abalation

To run MLM pretraining experiments (comparison of training regimes), cd to src/Training/MLM

CUDA_VISIBLE_DEVICES=0 python run_mlm_roberta.py --model_type roberta-mlm --model_name_or_path roberta-large --task_name atomicmlm --output_dir ../../out_dir --train_file \
../../data/ATOMIC/train_random.jsonl --dev_file ../../data/ATOMIC/dev_random.jsonl --mlm_probability 0.5 --max_seq_length 128 --max_words_to_mask 6 --max_sequence_per_time 200 \
--do_train --do_eval --per_gpu_train_batch_size 8 --gradient_accumulation_steps 4 --learning_rate 1e-5 --num_train_epochs 3 --warmup_proportion 0.05 --evaluate_during_training \
--per_gpu_eval_batch_size 8 --save_steps 5000

Then follow the same evaluation commands as above to evaluate the models

AFLite

To generate adversarial filtered datasets using AFLite algorithm, first run the data generation code with --do_split flag

python generate_from_ATOMIC.py --train_KG atomic/v4_atomic_trn.csv --dev_KG atomic/v4_atomic_dev.csv --strategy random --out_dir ../../data/ATOMIC --do_split 

This will split training set into 3 subsets, then we can train a feature function, cd to src/Training/AFLite directory

CUDA_VISIBLE_DEVICES=0 python run_roberta_classification.py --model_type roberta-mc --model_name_or_path roberta-large --task_name cwwv --output_dir ../../out_dir --train_file \
../../data/ATOMIC/train_4%_random.jsonl  --dev_file ../../data/ATOMIC/train_1%_random.jsonl --max_seq_length 128 --per_gpu_eval_batch_size 16 --do_train --do_eval \
--evaluate_during_training --per_gpu_train_batch_size 4 --gradient_accumulation_steps 8 --learning_rate 1e-5 --num_train_epochs 3 --warmup_proportion 0.05 --save_steps 150 

Then we compute the embeddings for the remaining 95% of train and dev sets

CUDA_VISIBLE_DEVICES=0 python run_roberta_classification.py --model_type roberta-mc --model_name_or_path roberta-large --task_name cwwv --output_dir ../../out_dir --train_file \
../../data/ATOMIC/train_4%_random.jsonl  --dev_file ../../data/ATOMIC/train_95%_random.jsonl --max_seq_length 128 --per_gpu_eval_batch_size 16 --do_eval
CUDA_VISIBLE_DEVICES=0 python run_roberta_classification.py --model_type roberta-mc --model_name_or_path roberta-large --task_name cwwv --output_dir ../../out_dir --train_file \
../../data/ATOMIC/train_4%_random.jsonl  --dev_file ../../data/ATOMIC/dev_random.jsonl --max_seq_length 128 --per_gpu_eval_batch_size 16 --do_eval

To run AFLite

python run_AFLite.py --train_file ../../data/ATOMIC/train_95%_random.jsonl  --dev_file ../../data/ATOMIC/dev_random.jsonl 

This will produce the AFLite filtered output files at the same location as input files, which can be used for pretraining the models.

Cite

@misc{ma2020knowledgedriven,
    title={Knowledge-driven Data Construction for Zero-shot Evaluation in Commonsense Question Answering},
    author={Kaixin Ma and Filip Ilievski and Jonathan Francis and Yonatan Bisk and Eric Nyberg and Alessandro Oltramari},
    year={2020},
    eprint={2011.03863},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}