Awesome

Birch

Document ranking via sentence modeling using BERT

Note: The results in the arXiv paper Simple Applications of BERT for Ad Hoc Document Retrieval have been superseded by the results in the EMNLP'19 paper [Cross-Domain Modeling of Sentence-Level Evidence for Document Retrieval]. To reproduce the results in the arXiv paper, please follow the instructions here instead.

Environment & Data

# Set up environment
pip install virtualenv
virtualenv -p python3.5 birch_env
source birch_env/bin/activate

# Install dependencies
pip install Cython  # jnius dependency
pip install -r requirements.txt

# For inference, the Python-only apex build can also be used
git clone https://github.com/NVIDIA/apex
cd apex && pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

# Set up Anserini (last reproduced with commit id: 5da46f610435be6364700bc5a6144253ed3f3b59)
git clone https://github.com/castorini/anserini.git
cd anserini && mvn clean package appassembler:assemble
cd eval && tar xvfz trec_eval.9.0.4.tar.gz && cd trec_eval.9.0.4 && make && cd ../../..

# Download data and models
wget https://zenodo.org/record/3381673/files/emnlp_bert4ir_v2.tar.gz
tar -xzvf emnlp_bert4ir_v2.tar.gz

Experiment Names:

• mb_robust04, mb_core17, mb_core18
• car_mb_robust04, car_mb_core17, car_mb_core18
• msmarco_mb_robust04, msmarco_mb_core17, msmarco_mb_core18
• robust04, car_core17, car_core18
• msmarco_robust04, msmarco_core17, msmarco_core18

Training

For BERT(MB):

export CUDA_VISIBLE_DEVICES=0; experiment=mb; \
nohup python -u src/main.py --mode training --experiment ${experiment} --collection mb \
--local_model models/bert-large-uncased.tar.gz \
--local_tokenizer models/bert-large-uncased-vocab.txt --batch_size 16 \
--data_path data --predict_path data/predictions/predict.${experiment} \
--model_path models/saved.${experiment} --eval_steps 1000 \
--device cuda --output_path logs/out.${experiment} > logs/${experiment}.log 2>&1 &

For BERT(CAR -> MB) and BERT(MS MARCO -> MB):

export CUDA_VISIBLE_DEVICES=0; experiment=<car_mb, msmarco_mb>; \
nohup python -u src/main.py --mode training --experiment ${experiment} --collection mb \
--local_model <models/pytorch_msmarco.tar.gz, models/pytorch_car.tar.gz> \
--local_tokenizer models/bert-large-uncased-vocab.txt --batch_size 16 \
--data_path data --predict_path data/predictions/predict.${experiment} \
--model_path models/saved.${experiment} --eval_steps 1000 \
--device cuda --output_path logs/out.${experiment} > logs/${experiment}.log 2>&1 &

Inference

For BERT(MB), BERT(CAR -> MB) and BERT(MS MARCO -> MB):

export CUDA_VISIBLE_DEVICES=0; experiment=<experiment_name>; \
nohup python -u src/main.py --mode inference --experiment ${experiment} --collection <robust04, core17, core18> \
--load_trained --model_path <models/saved.mb_1, models/saved.car_mb_1, models/saved.msmarco_mb_1> \
--batch_size 4 --data_path data --predict_path data/predictions/predict.${experiment} \
--device cuda --output_path logs/out.${experiment} > logs/${experiment}.log 2>&1 &

For BERT(CAR) and BERT(MS MARCO):

export CUDA_VISIBLE_DEVICES=0; experiment=<experiment_name; \
nohup python -u src/main.py --mode inference --experiment ${experiment} --collection <robust04, core17, core18> \
--local_model <models/pytorch_msmarco.tar.gz, models/pytorch_car.tar.gz> \
--local_tokenizer models/bert-large-uncased-vocab.txt --batch_size 4 \
--data_path data --predict_path data/predictions/predict.${experiment} \
--device cuda --output_path logs/out.${experiment} > logs/${experiment}.log 2>&1 &

Note that this step takes a long time. If you don't want to evaluate the pretrained models, you may skip to the next step and evaluate with our predictions under data/predictions.

Retrieve sentences from top candidate documents

python src/utils/split_docs.py --collection <robust04, core17, core18> \
--index <path/to/index> --data_path data --anserini_path <path/to/anserini/root>

Evaluation

experiment=<experiment_name>
collection=<robust04, core17, core18>
anserini_path=<path/to/anserini/root>
index_path=<path/to/lucene/index>
data_path=<path/to/data/root>

BM25+RM3 Baseline

./eval_scripts/baseline.sh ${collection} ${index_path} ${anserini_path} ${data_path}

./eval_scripts/eval.sh baseline ${collection} ${anserini_path} ${data_path}

Sentence Evidence

# Tune hyperparameters (if you do not have apex working, run this script with an additional "NOAPEX" param at the end)
./eval_scripts/train.sh ${experiment} ${collection} ${anserini_path}

# Run experiment
./eval_scripts/test.sh ${experiment} ${collection} ${anserini_path}

# Evaluate with trec_eval
./eval_scripts/eval.sh ${experiment} ${collection} ${anserini_path} ${data_path}