BPR
Binary Passage Retriever (BPR) is an efficient neural retrieval model for open-domain question answering. BPR integrates a learning-to-hash technique into Dense Passage Retriever (DPR) to represent the passage embeddings using compact binary codes rather than continuous vectors. It substantially reduces the memory size without a loss of accuracy tested on Natural Questions and TriviaQA datasets.
BPR was originally developed to improve the computational efficiency of the Sōseki question answering system submitted to the Systems under 6GB track in the NeurIPS 2020 EfficientQA competition. Please refer to our ACL 2021 paper for further technical details.
Installation
BPR can be installed using Poetry:
poetry install
The virtual environment automatically created by Poetry can be activated by poetry shell.
Alternatively, you can install required libraries using pip:
pip install -r requirements.txt
Trained Models
(coming soon)
Reproducing Experiments
Before you start, you need to download the datasets available on the DPR website into <DPR_DATASET_DIR>.
The experimental results on the Natural Questions dataset can be reproduced by running the commands provided in this section. We used a server with 8 NVIDIA Tesla V100 GPUs with 16GB memory in the experiments. The results on the TriviaQA dataset can be reproduced by changing the file names of the input dataset to the corresponding ones (e.g., nq-train.json -> trivia-train.json).
1. Building passage database
python build_passage_db.py \
--passage_file=<DPR_DATASET_DIR>/wikipedia_split/psgs_w100.tsv \
--output_file=<PASSAGE_DB_FILE>
2. Training BPR
python train_biencoder.py \
--gpus=8 \
--distributed_backend=ddp \
--train_file=<DPR_DATASET_DIR>/retriever/nq-train.json \
--eval_file=<DPR_DATASET_DIR>/retriever/nq-dev.json \
--gradient_clip_val=2.0 \
--max_epochs=40 \
--binary
3. Building passage embeddings
python generate_embeddings.py \
--biencoder_file=<BPR_CHECKPOINT_FILE> \
--output_file=<EMBEDDING_FILE> \
--passage_db_file=<PASSAGE_DB_FILE> \
--batch_size=4096 \
--parallel
4. Evaluating BPR
python evaluate_retriever.py \
--binary_k=1000 \
--biencoder_file=<BPR_CHECKPOINT_FILE> \
--embedding_file=<EMBEDDING_FILE> \
--passage_db_file=<PASSAGE_DB_FILE> \
--qa_file=<DPR_DATASET_DIR>/retriever/qas/nq-test.csv \
--parallel
5. Creating dataset for reader
python evaluate_retriever.py \
--binary_k=1000 \
--biencoder_file=<BPR_CHECKPOINT_FILE> \
--embedding_file=<EMBEDDING_FILE> \
--passage_db_file=<PASSAGE_DB_FILE> \
--qa_file=<DPR_DATASET_DIR>/retriever/qas/nq-train.csv \
--output_file=<READER_TRAIN_FILE> \
--top_k=200 \
--parallel
python evaluate_retriever.py \
--binary_k=1000 \
--biencoder_file=<BPR_CHECKPOINT_FILE> \
--embedding_file=<EMBEDDING_FILE> \
--passage_db_file=<PASSAGE_DB_FILE> \
--qa_file=<DPR_DATASET_DIR>/retriever/qas/nq-dev.csv \
--output_file=<READER_DEV_FILE> \
--top_k=200 \
--parallel
python evaluate_retriever.py \
--binary_k=1000 \
--biencoder_file=<BPR_CHECKPOINT_FILE> \
--embedding_file=<EMBEDDING_FILE> \
--passage_db_file=<PASSAGE_DB_FILE> \
--qa_file==<DPR_DATASET_DIR>/retriever/qas/nq-test.csv \
--output_file=<READER_TEST_FILE> \
--top_k=200 \
--parallel
6. Training reader
python train_reader.py \
--gpus=8 \
--distributed_backend=ddp \
--train_file=<READER_TRAIN_FILE> \
--validation_file=<READER_DEV_FILE> \
--test_file=<READER_TEST_FILE> \
--learning_rate=2e-5 \
--max_epochs=20 \
--accumulate_grad_batches=4 \
--nq_gold_train_file=<DPR_DATASET_DIR>/gold_passages_info/nq_train.json \
--nq_gold_validation_file=<DPR_DATASET_DIR>/gold_passages_info/nq_dev.json \
--nq_gold_test_file=<DPR_DATASET_DIR>/gold_passages_info/nq_test.json \
--train_batch_size=1 \
--eval_batch_size=2 \
--gradient_clip_val=2.0
7. Evaluating reader
python evaluate_reader.py \
--gpus=8 \
--distributed_backend=ddp \
--checkpoint_file=<READER_CHECKPOINT_FILE> \
--eval_batch_size=1
License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Citation
If you find this work useful, please cite the following paper:
@inproceedings{yamada2021bpr,
title={Efficient Passage Retrieval with Hashing for Open-domain Question Answering},
author={Ikuya Yamada and Akari Asai and Hannaneh Hajishirzi},
booktitle={ACL},
year={2021}
}
70 Nov 27, 2022
34 Apr 13, 2022
22 Nov 9, 2022
47 Jul 26, 2022
434 Jan 4, 2023
76 Dec 21, 2022
19 Sep 9, 2021
50 Nov 24, 2022
40 Dec 13, 2022
310 Dec 28, 2022
33 Dec 5, 2022
506 Nov 29, 2022
41 Nov 8, 2022
672 Jan 1, 2023
18 Dec 9, 2022
72 Dec 5, 2022
68 Jul 18, 2022
49 Nov 10, 2022