Introduction

We have published several works on generative retrieval as follows.

Multiview Identifiers Enhanced Generative Retrieval. ACL 2023. (MINDER)
Generative Retrieval for Conversational Question Answering. IPM 2023. (GCoQA)
Learning to Rank in Generative Retrieval. AAAI 2024. (LTRGR)
Generative Cross-Modal Retrieval: Memorizing Images in Multimodal Language Models for Retrieval and Beyond. ACL 2024 (GRACE).
Distillation Enhanced Generative Retrieval. ACL 2024 findings (DGR).

All code, data, and checkpoints of the above works are open-released:

1. MINDER, LTRGR, and DGR, are a series of works on text retrieval. LTRGR and DGR are continuously training based on the MINDER model, so we release MINDER, LTRGR, and DGR together in the same repository https://github.com/liyongqi67/MINDER.
2. GCoQA is the work on conversational retrieval and is released at https://github.com/liyongqi67/GCoQA.
3. GRACE is the work on cross-modal retrieval and is released at https://github.com/liyongqi67/GRACE.

You could also refer to our preprint works on generative retrieval.

A Survey of Generative Search and Recommendation in the Era of Large Language Models.
Revolutionizing Text-to-Image Retrieval as Autoregressive Token-to-Voken Generation.

MINDER

This is the official implementation for the paper "Multiview Identifiers Enhanced Generative Retrieval".
The preprint version is released in Arxiv.
If you find our paper or code helpful, please consider citing as follows:

@inproceedings{li-etal-2023-multiview,
    title = "Multiview Identifiers Enhanced Generative Retrieval",
    author = "Li, Yongqi  and Yang, Nan  and Wang, Liang  and Wei, Furu  and Li, Wenjie",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    publisher = "Association for Computational Linguistics",
    pages = "6636--6648",
}

Install

git clone https://github.com/liyongqi67/MINDER.git
sudo apt install swig
env CFLAGS='-fPIC' CXXFLAGS='-fPIC' res/external/sdsl-lite/install.sh
pip install -r requirements.txt
pip install -e .

Data

Option 1: Please download the processed data into the data folder.

1. data/NQ folder. Please download biencoder-nq-dev.json, biencoder-nq-train.json, nq-dev.qa.csv, nq-test.qa.csv files into the NQ folder from the DPR repositiory.
2. data/Trivia folder. Please download biencoder-trivia-dev.json, biencoder-trivia-train.json, trivia-dev.qa.csv, trivia-test.qa.csv files into the Trivia folder from the DPR repositiory.
3. data/MSMARCO folder. Please download qrels.msmarco-passage.dev-subset.txt from this link.
4. data/fm_index/ folder. Please download fm_index files psgs_w100.fm_index.fmi, psgs_w100.fm_index.oth for the Wikipedia corpus and msmarco-passage-corpus.fm_index.fmi, msmarco-passage-corpus.fm_index.oth for the MSMARCO corpus from this link.
5. data/training_data/ folder.
   Download the NQ_title_body_query_generated from this link.
   Download the Trivia_title_body_query_generated from this link.
   Download the MSMARCO_title_body_query3 from this link.
6. data/pseudo_queries/ folder. link pseudo_queries for Wikipedia and MSMARCO.

Option 2: You could process the data by yourself following the instructions.

Model training

We use the fairseq to train the BART_large model with the translation task.
The script for training on the NQ dataset is

    - fairseq-train
        data/training_data/NQ_title_body_query_generated/bin 
        --finetune-from-model /bart.large/model.pt 
        --arch bart_large 
        --task translation 
        --criterion label_smoothed_cross_entropy 
        --source-lang source 
        --target-lang target 
        --truncate-source 
        --label-smoothing 0.1 
        --max-tokens 4096 
        --update-freq 1 
        --max-update 800000 
        --required-batch-size-multiple 1
        --validate-interval 1000000
        --save-interval 1000000
        --save-interval-updates 15000 
        --keep-interval-updates 3 
        --dropout 0.1 
        --attention-dropout 0.1 
        --relu-dropout 0.0 
        --weight-decay 0.01 
        --optimizer adam 
        --adam-betas "(0.9, 0.999)" 
        --adam-eps 1e-08 
        --clip-norm 0.1 
        --lr-scheduler polynomial_decay 
        --lr 3e-05 
        --total-num-update 800000 
        --warmup-updates 500 
        --fp16 
        --num-workers 10 
        --no-epoch-checkpoints 
        --share-all-embeddings 
        --layernorm-embedding 
        --share-decoder-input-output-embed 
        --skip-invalid-size-inputs-valid-test 
        --log-format json
        --log-interval 100 
        --patience 5
        --find-unused-parameters
        --save-dir  ./

The script for training on the TriviaQA dataset is

    - fairseq-train
        data/training_data/Trivia_title_body_query_generated/bin 
        --finetune-from-model /bart.large/model.pt 
        --arch bart_large 
        --task translation 
        --criterion label_smoothed_cross_entropy 
        --source-lang source 
        --target-lang target 
        --truncate-source 
        --label-smoothing 0.1 
        --max-tokens 4096 
        --update-freq 1 
        --max-update 800000 
        --required-batch-size-multiple 1
        --validate-interval 1000000
        --save-interval 1000000
        --save-interval-updates 6000 
        --keep-interval-updates 3 
        --dropout 0.1 
        --attention-dropout 0.1 
        --relu-dropout 0.0 
        --weight-decay 0.01 
        --optimizer adam 
        --adam-betas "(0.9, 0.999)" 
        --adam-eps 1e-08 
        --clip-norm 0.1 
        --lr-scheduler polynomial_decay 
        --lr 3e-05 
        --total-num-update 800000 
        --warmup-updates 500 
        --fp16 
        --num-workers 10 
        --no-epoch-checkpoints 
        --share-all-embeddings 
        --layernorm-embedding 
        --share-decoder-input-output-embed 
        --skip-invalid-size-inputs-valid-test 
        --log-format json
        --log-interval 100 
        --patience 5
        --find-unused-parameters
        --save-dir  ./

The script for training on the MSMARCO dataset is

    - fairseq-train
        data/training_data/MSMARCO_title_body_query3/bin 
        --finetune-from-model /bart.large/model.pt 
        --arch bart_large 
        --task translation 
        --criterion label_smoothed_cross_entropy 
        --source-lang source 
        --target-lang target 
        --truncate-source 
        --label-smoothing 0.1 
        --max-tokens 4096 
        --update-freq 1 
        --max-update 100000 
        --required-batch-size-multiple 1
        --validate-interval 1000000
        --save-interval 1000000
        --save-interval-updates 6000 
        --keep-interval-updates 3 
        --dropout 0.1 
        --attention-dropout 0.1 
        --relu-dropout 0.0 
        --weight-decay 0.01 
        --optimizer adam 
        --adam-betas "(0.9, 0.999)" 
        --adam-eps 1e-08 
        --clip-norm 0.1 
        --lr-scheduler polynomial_decay 
        --lr 3e-05 
        --total-num-update 100000 
        --warmup-updates 500 
        --fp16 
        --num-workers 10 
        --no-epoch-checkpoints 
        --share-all-embeddings 
        --layernorm-embedding 
        --share-decoder-input-output-embed 
        --skip-invalid-size-inputs-valid-test 
        --log-format json
        --log-interval 100 
        --patience 3
        --find-unused-parameters
        --save-dir  ./

We trained the models on 8*32GB NVIDIA V100 GPUs. It took about 4d3h24m39s, 1d18h30m47s, 12h53m50s for training on NQ, TriviaQA, and MSMARCO, respectively.
We release our trained model checkpoints in this link.

Model inference

Please use the following script to retrieve passages for queries in NQ.

    - TOKENIZERS_PARALLELISM=false python seal/search.py 
      --topics_format dpr_qas --topics data/NQ/nq-test.qa.csv 
      --output_format dpr --output output_test.json 
      --checkpoint checkpoint_NQ.pt 
      --jobs 10 --progress --device cuda:0 --batch_size 20 
      --beam 15
      --decode_query stable
      --fm_index data/fm_index/stable2/psgs_w100.fm_index 

Please use the following script to retrieve passages for queries in TriviaQA.

    - TOKENIZERS_PARALLELISM=false python seal/search.py 
      --topics_format dpr_qas --topics data/Trivia/trivia-test.qa.csv
      --output_format dpr --output output_test.json 
      --checkpoint checkpoint_TriviaQA.pt 
      --jobs 10 --progress --device cuda:0 --batch_size 40 
      --beam 15
      --decode_query stable
      --fm_index data/fm_index/stable2/psgs_w100.fm_index

Please use the following script to retrieve passages for queries in MSMARCO.

    - TOKENIZERS_PARALLELISM=false python seal/search.py 
      --topics_format msmarco --topics Tevatron/msmarco-passage
      --output_format msmarco --output output_test.json 
      --checkpoint checkpoint_MSMARCO.pt 
      --jobs 10 --progress --device cuda:0 --batch_size 10 
      --beam 7
      --decode_query stable
      --fm_index data/fm_index/stable2/msmarco-passage-corpus.fm_index

Evaluation

Please use the following script to evaluate on NQ and TriviaQA.

    - python3 seal/evaluate_output.py
      --file output_test.json 

Please use the following script to evaluate on MSMARCO.

    - python3 seal/evaluate_output_msmarco.py
      data/MSMARCO/qrels.msmarco-passage.dev-subset.txt output_test.json

LTRGR

This is the official implementation for the paper "Learning to Rank in Generative Retrieval".
The preprint version is released in Arxiv. If you find our paper or code helpful,please consider citing as follows:

@article{li2023learning,
  title={Learning to rank in generative retrieval},
  author={Li, Yongqi and Yang, Nan and Wang, Liang and Wei, Furu and Li, Wenjie},
  journal={arXiv preprint arXiv:2306.15222},
  year={2023}
}

Install

git clone https://github.com/liyongqi67/MINDER.git
sudo apt install swig
env CFLAGS='-fPIC' CXXFLAGS='-fPIC' res/external/sdsl-lite/install.sh
pip install -r requirements.txt
pip install -e .

Model Training

You could directly download our trained checkpoints.

Learning to generate

Learning to generate means to train the MINDER. You could refer to the above MINDER training procedures or load the trained MINDER checkpoints.

Learning to rank

Step 1: Data preparation. Load the MINDER checkpoints and obtain the top-200 retrieval results on the training set.
On NQ

    TOKENIZERS_PARALLELISM=false python seal/search.py 
      --topics_format dpr_qas_train --topics data/NQ/biencoder-nq-train.json 
      --output_format dpr --output MINDER_NQ_train_top200.json 
      --checkpoint checkpoint_NQ.pt 
      --jobs 10 --progress --device cuda:0 --batch_size 10 
      --beam 15
      --decode_query stable
      --fm_index data/fm_index/stable2/psgs_w100.fm_index
      --include_keys
      --hits 200

Step 2: Train the model via rank loss.

    TOKENIZERS_PARALLELISM=false  python3 MINDER_learn_to_rank_v1.4.py 
        --checkpoint checkpoint_NQ.pt
        --fm_index data/fm_index/stable2/psgs_w100.fm_index
        --train_file MINDER_NQ_train_top200.json
        --do_fm_index True
        --per_gpu_train_batch_size 8
        --output_dir ./release_test
        --rescore_batch_size 70
        --num_train_epochs 3
        --factor_of_generation_loss 1000
        --rank_margin 300
        --shuffle_positives True
        --shuffle_negatives True
        --shuffle_negatives True
        --decode_query stable
        --pid2query pid2query_Wikipedia.pkl

Model inference

Please use the following script to retrieve passages for queries in NQ.

for file in ./release_test/*
do
    if test -f $file
    then
        echo $file
        TOKENIZERS_PARALLELISM=false python -m seal.search \
        --topics_format dpr_qas --topics data/NQ/nq-test.qa.csv \
        --output_format dpr --output output_test.json \
        --checkpoint $file\
        --jobs 5 --progress --device cuda:0 --batch_size 10 --beam 15 \
        --decode_query stable --fm_index data/fm_index/stable2/psgs_w100.fm_index --dont_fairseq_checkpoint
        python3 evaluate_output.py --file output_test.json
    fi
done

DGR

This is the official implementation for the paper "Distillation Enhanced Generative Retrieval". The preprint version is released in Arxiv. If you find our paper or code helpful,please consider citing as follows:

@misc{li2024distillation,
      title={Distillation Enhanced Generative Retrieval}, 
      author={Yongqi Li and Zhen Zhang and Wenjie Wang and Liqiang Nie and Wenjie Li and Tat-Seng Chua},
      year={2024},
      eprint={2402.10769},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Install

git clone https://github.com/liyongqi67/MINDER.git
sudo apt install swig
env CFLAGS='-fPIC' CXXFLAGS='-fPIC' res/external/sdsl-lite/install.sh
pip install -r requirements_dgr.txt
pip install -e .

Model Training

You could directly download our trained DGR checkpoints

Learning to generate

Learning to generate means to train the MINDER. You could refer to the above MINDER training procedures or load the trained MINDER checkpoints.

Distillation enhanced generative retrieval training

Get training data

Step 1: Data preparation. Load the MINDER checkpoints and obtain the top-200 retrieval results on the training set.
Take NQ as an example. For TriviaQA and MsMarco, you may change data and checkpoints, which can be download from MINDER.

TOKENIZERS_PARALLELISM=false python seal/search.py 
    --topics_format dpr_qas_train --topics data/NQ/biencoder-nq-train.json 
    --output_format dpr --output ./TrainData/NQ/MINDER_NQ_train_top200.json 
    --checkpoint checkpoint_NQ.pt 
    --jobs 10 --progress --device cuda:0 --batch_size 10 
    --beam 15
    --decode_query stable
    --fm_index data/fm_index/stable2/psgs_w100.fm_index
    --include_keys
    --hits 200

Step 2: Obtain ranking score of teacher model for the training question-answer pairs. We use E5 and SimLM as ranking teacher. On NQ and SimLM as teacher(use simLM specially trained on NQ data), which can be download DGR checkpoints.

python ./seal/check_top200.py  
--teacher_output_path ./TrainData/simLM_NQ 
--original_path ./TrainData/NQ  
--do_train Ture --teacher_path ./nq-simLM-reranker/reranker-model
--teacher_name simLM --teacher_kind score_teacher_simLM_NQ 
--original_file MINDER_NQ_train_top200.json

On NQ and E5 as teacher

python ./seal/check_top200.py  
--teacher_output_path ./TrainData/E5_NQ 
--original_path ./TrainData/NQ
--do_train Ture --teacher_path intfloat/e5-large-v2
--teacher_name E5 --teacher_kind score_teacher_E5
--original_file MINDER_NQ_train_top200.json

On TriviaQA and SimLM as teacher(use simLM specially trained on NQ data)

python ./seal/check_top200.py 
--teacher_output_path ./TrainData/simLM_Trival_QA  
--original_path ./Train_data_Triva_QA --teacher_path ./nq-simLM-reranker/reranker-model
--teacher_name simLM --teacher_kind score_teacher_simLM_NQ 
--original_file MINDER_triviaQA_train_top200.json

On TriviaQA and E5 as teacher

python ./seal/check_top200.py 
--teacher_output_path ./TrainData/simLM_Trival_QA  
--do_train Ture --teacher_path intfloat/e5-large-v2
--teacher_name E5 --teacher_kind score_teacher_E5
--original_file MINDER_triviaQA_train_top200.json

On MsMarco and simLM as teacher

python ./seal/check_top200.py  
--teacher_output_path ./TrainData/simLM_MSMarco 
--original_path ./TrainData/MSMarco --do_train Ture --teacher_path intfloat/simlm-msmarco-reranker 
--teacher_name simLM_msmarco --teacher_kind score_teacher_simLM_msmarco 
--if_MSMarco True

Distillation enhanced generative retrieval training process

train the model via distilled RankNet loss. On NQ and simLM as teacher

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --standalone --nnodes=1 --nproc-per-node=4 ./seal/minder_ltr_listwise.py --decode_query stable --train_file ./TrainData/simLM_NQ/ --per_gpu_train_batch_size 1  --factor_of_generation_loss 500 --shuffle_positives True --shuffle_negatives True --do_fm_index True --sort_negative_only True --order_margin True --rest_all_negative False --sample_length 6 --num_train_epochs 3 --rank_margin 300 --teacher_kind score_teacher_simLM_NQ --manual_rankNet True

For training on NQ based on E5 as teacher, we only need to change 'teacher_kind' to 'score_teacher_E5'

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --standalone --nnodes=1 --nproc-per-node=4 ./seal/minder_ltr_listwise.py --decode_query stable --train_file ./TrainData/E5_NQ/ --per_gpu_train_batch_size 1  --factor_of_generation_loss 500 --shuffle_positives True --shuffle_negatives True --do_fm_index True --sort_negative_only True --order_margin True --rest_all_negative False --sample_length 6 --num_train_epochs 3 --rank_margin 300 --teacher_kind score_teacher_E5 --manual_rankNet True

For TriviaQA, we only need to change training data 'train_file' to './TrainData/simLM_Trival_QA/' and teacher_kind to 'score_teacher_simLM_NQ'

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --standalone --nnodes=1 --nproc-per-node=4 ./seal/minder_ltr_listwise.py --decode_query stable --train_file ./TrainData/simLM_Trival_QA/ --per_gpu_train_batch_size 1  --factor_of_generation_loss 500 --shuffle_positives True --shuffle_negatives True --do_fm_index True --sort_negative_only True --order_margin True --rest_all_negative False --sample_length 6 --num_train_epochs 3 --rank_margin 300 --teacher_kind score_teacher_simLM_NQ --manual_rankNet True

For MsMarco, we only need to change training data 'train_file' to './TrainData/simLM_MSMarco/', and change 'teacher_kind' to 'score_teacher_simLM_msmarco'

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --standalone --nnodes=1 --nproc-per-node=4 ./seal/minder_ltr_listwise.py --decode_query stable --train_file ./TrainData/MSMarco_train_all_simLM_NQ_marco/ --per_gpu_train_batch_size 1  --factor_of_generation_loss 500 --shuffle_positives True --shuffle_negatives True --do_fm_index True --sort_negative_only True --order_margin True --rest_all_negative False --sample_length 6 --num_train_epochs 3 --rank_margin 300 --teacher_kind score_teacher_simLM_msmarco --manual_rankNet True --checkpoint ./checkpoint_MSMARCO.pt  --fm_index ./data/fm_index/msmarco-passage-corpus.fm_index
--pid2query ./data/pid2query_msmarco.pkl

To train model with learning to rank loss such as listNet loss, ListMLE, LambdaLoss, you can clone the repo allRank into the root directory of this project. Then run

PWD = `pwd`
export PYTHONPATH=$PWD:$PWD/allRank:$PYTHONPATH
CUDA_VISIBLE_DEVICES=0,1,2,3  torchrun --standalone --nnodes=1 --nproc-per-node=4 ./seal/minder_ltr_listwise.py --decode_query stable --train_file ./TrainData/simLM_NQ/  --per_gpu_train_batch_size 1 --experiment_name SimLM_NQ_lbdloss_weight3_list2_m300 --scenario_name   updated_two_posi_sort_nega_only_anlysis  --shuffle_positives True --shuffle_negatives True  --do_fm_index True --sort_negative_only True --order_margin True --rest_all_negative False --sample_length 6 --num_train_epochs 3 --teacher_kind score_teacher_simLM_NQ --factor_of_generation_loss 1 --margin_use_ltrlloss True --loss_func lambdaLoss

Model Inference

Please use the following script to retrieve passages for queries in NQ.

for file in ./release_test/*
do
    if test -f $file
    then
        echo $file
        TOKENIZERS_PARALLELISM=false python -m seal.search \
        --topics_format dpr_qas --topics data/NQ/nq-test.qa.csv \
        --output_format dpr --output output_test.json \
        --checkpoint $file\
        --jobs 5 --progress --device cuda:0 --batch_size 10 --beam 15 \
        --decode_query stable --fm_index data/fm_index/stable2/psgs_w100.fm_index --dont_fairseq_checkpoint
        python3 evaluate_output.py --file output_test.json
    fi
done

Acknowledgments

Part of the code is based on SEAL and sdsl-lite.

Contact

If there is any problem, please email liyongqi0@gmail.com. Please do not hesitate to email me directly as I do not frequently check GitHub issues.