It is based on BERT model.
We use "BERT-Base, Multilingual Cased" model for Korean.
We share training/test data, script, and a little modified source code in this repository.
You can download a BERT and MODEL, then easily combine to adjust a Korean RE model.
Train: sentence_classifier.sh
BERT_BASE_DIR: BERT Language Model, multi_cased_L-12_H-768_A-12
KORE_DIR: Dataset path
TRAINED_CLASSIFIER: Path to trained model (save)
--do_lower_case=False \ # boolean, lower case or not in BERT language model
--task_name=STFL \ # String, task name (KORE: Korean Relation Extraction)
--do_train=true \ # boolean, train or not
--do_eval=true \ # boolean, evaluation or not
--data_dir=$KORE_DIR/ \ # String, data file (train, dev, test) path
--vocab_file=$BERT_BASE_DIR/vocab.txt \ # String, BERT language model vocabulary path
--bert_config_file=$BERT_BASE_DIR/bert_config.json \ # String, BERT config file path
--init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \ # String, BERT model check point path
--max_seq_length=128 \ # Integer, max input sequence length
--train_batch_size=16 \ # Integer, batch size
--learning_rate=2e-5 \ # Double, learning rate
--num_train_epochs=30.0 \ # Integer, train epoches
--output_dir=trained_model/sentence_filtering/ \ # String, Path to trained model (save)
--output_csv=ko_re_data/output.csv # String, Path to evaluation result
Test: sentence_prediction.sh
preproc_filter.py [INPUT_CSV] [OUTPUT_TSV]
and
Set train/dev/test file path in class StflProcessor - run_classifier.py
and
sentence_classifier.sh or sentence_prediction.sh
and
postproc_NA.py
and
Use ko_re_data/filtered_sentence_True.csv file
NA 썰전 대한민국 종합편성채널 JTBC ‘<e1>썰전</e1> ’이 예능심판자 코너를 통해 아역배우 김동현_(1984년) 군과 함께 <e2>대한민국</e2> 아역스타들을 분석하는 시간을 가졌다.
TSV format; relation, subject, object, and sentence with two entities
and
"구한말에는 갑신정변 으로 피신했다가 귀국, <e1>독립협회</e1> 활동, 독립신문 <e2>발행인</e2> 과 제2대 독립신문 (獨立新聞社) 사장 등으로 활동했으며 만민공동회 의 최고 지도력 로서 강연, 계몽활동과 민권운동과 민중의 참정권 요구 운동·개혁운동에 참여했고, 서재필 이 강제추방된 이후 <e1>독립협회</e1> 와 반청계몽운동 활동을 지도했으나, 후에 친일 인사로 변절하였다.",NA,,,,0.8557533
<e1>GRP_레코드</e1> ()는 <e2>미국</e2> 의 레코드 회사이다.,True
Output CSV format; sentence with two entities, relation (NA or True), score
CC BY-NC-SAAttribution-NonCommercial-ShareAlike- If you want to commercialize this resource, please contact to us
Machine Reading Lab @ KAIST
This work was supported by Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (2013-0-00109, WiseKB: Big data based self-evolving knowledge base and reasoning platform)