Skip to content

Latest commit

 

History

History
172 lines (142 loc) · 4.42 KB

LAUNCH.md

File metadata and controls

172 lines (142 loc) · 4.42 KB
Raw

DrFact

Installation

# Note that we use TF 1.15. This is because we use the tf.contrib package,
# which was removed from TF 2.0.
conda create --name drfact python=3.7
conda activate drfact
pip install tensorflow==1.15
# pip install tensorflow-gpu==1.15.0
# We use TensorFlow Hub to download BERT, which is used to initialize DrFact.
pip install tensorflow-hub bert-tensorflow
pip install spacy -U
python -m spacy download en_core_web_lg

Corpus Preprocessing

Download the GenericsKB corpus.

Link: https://allenai.org/data/genericskb

Put the GenericsKB-Best.tsv file under the ./knowledge_corpus/ folder. Note that the language and knowledge_corpus folder are at the same level.

Run CorpusPrepro.ipynb, NounChunkVocab.ipynb and ChangeFormat.ipynb. All the preprocessed data here should be at /path/to/knowledge_corpus/.

Indexing

Put the BERT ckpts bert/pretrained_models/wwm_uncased_L-24_H-1024_A-16/* to /path/to/bert_large/.

Set up the environment.

CORPUS_PATH=/path/to/knowledge_corpus/
BERT_PATH=/path/to/bert_large/
INDEX_PATH=/path/to/local_drfact_index

Basic data preprocessing

python index_corpus.py \
--do_preprocess \
--entity_file ${CORPUS_PATH}/gkb_best.vocab.txt \
--wiki_file ${CORPUS_PATH}/gkb_best.drkit_format.jsonl \
--max_entity_length 5 \
--max_mentions_per_doc 20 \
--tokenizer_type bert_tokenization \
--vocab_file ${BERT_PATH}/vocab.txt \
--multihop_output_dir ${INDEX_PATH}/drfact_output_bert200 \
--alsologtostderr

Fact2Fact matrix computation

Run indexing in parallel (num_shards=3).

for (( c=0; c<=3; c++ ))
do
   python fact2fact_index.py \
    --do_preprocess \
    --entity_file ${CORPUS_PATH}/gkb_best.vocab.txt \
    --wiki_file ${CORPUS_PATH}/gkb_best.drkit_format.jsonl \
    --fact2fact_index_dir ${INDEX_PATH}/fact2fact_index \
    --multihop_output_dir ${INDEX_PATH}/drfact_output_bert200 \
    --num_shards 3 --my_shard $c \
    --alsologtostderr
done

Combine the single files.

python fact2fact_index.py \
  --do_combine \
  --entity_file ${CORPUS_PATH}/gkb_best.vocab.txt \
  --wiki_file ${CORPUS_PATH}/gkb_best.drkit_format.jsonl \
  --fact2fact_index_dir ${INDEX_PATH}/fact2fact_index \
  --multihop_output_dir ${INDEX_PATH}/drfact_output_bert200 \
  --num_shards 3 \
  --alsologtostderr

Embedding the facts with BERT.

(Note that we can also embed mentions by replace do_embed_facts with do_embed_mentions).

Run computation (num_shards=1 if only have one gpu).

python index_corpus.py \
--do_embed \
--do_embed_facts \
--entity_file ${CORPUS_PATH}/gkb_best.vocab.txt \
--wiki_file ${CORPUS_PATH}/gkb_best.drkit_format.jsonl \
--max_entity_length 5 \
--max_mentions_per_doc 20 \
--tokenizer_type bert_tokenization \
--bert_ckpt_dir ${BERT_PATH}/ \
--vocab_file ${BERT_PATH}/vocab.txt \
--bert_config_file ${BERT_PATH}/bert_config.json \
--ckpt_name bert_model.ckpt --embed_prefix bert_large \
--multihop_output_dir ${INDEX_PATH}/drfact_output_bert200 \
--predict_batch_size 128 \
--max_seq_length 128 \
--doc_layers_to_use -1 \
--doc_aggregation_fn concat \
--qry_layers_to_use 4 \
--qry_aggregation_fn concat \
--max_query_length  64 \
--projection_dim 200 \
--doc_stride 128 \
--num_shards 1 \
--alsologtostderr

Run combination.

python index_corpus.py \
--do_combine \
--do_embed_facts \
--max_entity_length 5 \
--max_mentions_per_doc 20 \
--multihop_output_dir ${INDEX_PATH}/drfact_output_bert200 \
--num_shards 1 \
--projection_dim 200 \
--tokenizer_type dummy \
--shards_to_combine 1 \
--alsologtostderr

Clean the data:

PREFIX=bert_large
rm ${INDEX_PATH}/drfact_output_bert200/${PREFIX}_fact_feats_*.*

Dataset Processing

Run the QADataPrepro.ipynb for preprocessing the datasets.

Then, we link the concepts in the questions and answers:

ORI_DATA_DIR=/path/to/datasets/ARC-Easy/;
VOCAB=${CORPUS_PATH}/gkb_best.vocab.txt

SPLIT=arc_easy_train NUM_CHOICES=-1;  \
python link_questions.py \
  --csqa_file $ORI_DATA_DIR/${SPLIT}_processed.jsonl \
  --output_file $DATA_DIR/linked_${SPLIT}.all.jsonl \
  --indexed_concept_file  ${VOCAB} \
  --do_filtering ${NUM_CHOICES} \
  --alsologtostderr

SPLIT=arc_easy_dev NUM_CHOICES=3; \
python link_questions.py \
  --csqa_file $ORI_DATA_DIR/${SPLIT}_processed.jsonl \
  --output_file $DATA_DIR/linked_${SPLIT}.all.jsonl \
  --indexed_concept_file ${VOCAB} \
  --do_filtering ${NUM_CHOICES} \
  --alsologtostderr

Data Analysis

Model Training and Evaluation

bash drfact.sh