kaimss / anonymous_kowow

This repository forked from parlAI. Korean Wizard of Wikipedia task was added to this repo. This repository is going to be moved after EMNLP blind review.

Install parlai package

python setup.py develop

Download Korean Wizard of Wikipedia (KoWoW)

Download

tar xvzf wizard_of_wikipedia_ko.tar.gz
mv wizard_of_wikipedia_ko data/

language combinations

Name	Knowledge	Utterance
ko	Korean	Korean
ke	Korean	English
ek	English	Korean
en	English	English

Test splits

Name	topic
random_split	Seen
topic_split	Unseen

training on KoWoW (ek dataset)

CUDA_VIS_DEV=0
LANG_TYPE=ek

NUM_EPOCHS=10
MODEL_TYPE=T5EndToEndTwoAgent
MODEL_NAME=mymodel

CUDA_VISIBLE_DEVICES="${CUDA_VIS_DEV}" parlai train_model -t wizard_of_wikipedia_ko:generator:topic_split --ln ${LANG_TYPE} -m projects.wizard_of_wikipedia_ko.generator.t5:${MODEL_TYPE}  -mf model_data/model/${MODEL_NAME}  --t5-model-arch "KETI-AIR/ke-t5-base" --t5-encoder-model-arch "KETI-AIR/ke-t5-small" --log-every-n-secs 10 --validation-patience 12 --validation-metric ppl --validation-metric-mode min --validation-every-n-epochs 5 -bs 4 --max_knowledge 32 --num-epochs ${NUM_EPOCHS} --tensorboard-log true --tensorboard-logdir ./model_data/tf_logs/${MODEL_NAME}

test topic split(Unseen) on KoWoW ek dataset

CUDA_VIS_DEV=0
LANG_TYPE=ek
MODEL_NAME=mymodel

CUDA_VISIBLE_DEVICES="${CUDA_VIS_DEV}" parlai eval_model -t wizard_of_wikipedia_ko:generator:topic_split -dt test --ln ${LANG_TYPE} -mf model_data/model/${MODEL_NAME} -bs 4 --inference beam --beam-size 4

Display sample - topic split(Unseen)

You can display predicted knowledge using enc_output field.

CUDA_VIS_DEV=0
LANG_TYPE=ek
MODEL_NAME=mymodel
NUM_EXAMPLES=300

CUDA_VISIBLE_DEVICES="${CUDA_VIS_DEV}" parlai display_model -t wizard_of_wikipedia_ko:generator:topic_split -dt test --ln ${LANG_TYPE} -mf ./model_data/model/${MODEL_NAME} -bs 1 --inference beam --beam-size 4 --display-add-fields checked_sentence,enc_output -n NUM_EXAMPLES