Skip to content

Latest commit

 

History

History
253 lines (245 loc) · 10.7 KB

README.md

File metadata and controls

253 lines (245 loc) · 10.7 KB
Raw

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.

GCoQA

This is the official implementation for the paper "Generative Retrieval for Conversational Question Answering".
The paper is released in link.
If you find our paper or code helpful,please consider citing as follows:

@article{LI2023103475,
title = {Generative retrieval for conversational question answering},
journal = {Information Processing & Management},
volume = {60},
number = {5},
pages = {103475},
year = {2023},
issn = {0306-4573},
doi = {https://doi.org/10.1016/j.ipm.2023.103475},
url = {https://www.sciencedirect.com/science/article/pii/S0306457323002121},
author = {Yongqi Li and Nan Yang and Liang Wang and Furu Wei and Wenjie Li},
}

Dataset

We conducted experiments on three conversational open-domain QA datasets: OR-QuAC, QRECC, and TOPIOCQA. To facilitate future research in this area, we unified the three datasets into a benchmark with the same corpus, as DPR did.

1. Corpus.

1.1 Passage-level corpus: full_wiki_segments.json.
Format:

{
'id': 0,
'title': 'Eliza Fletcher [SEP] Introduction',
'text': 'Eliza Fletcher, née Dawson (15 January 1770 – 5 February 1858) was an English autobiographer and early travel writer.'
}

"Eliza Fletcher" is the page title, and "Introduction" is the section title, in Wikipedia.
1.2 Document-level corpus: full_wiki_document.json
Format:

{
'id': 0,
'title': 'Eliza Fletcher',
'text': '......'
}

"Eliza Fletcher" is the page title in Wikipedia.

2. QA pairs.

TOPIOCQA dataset: topiocqa_train.json, topiocqa_dev.json, topiocqa_test.json.
QRECC dataset: qrecc_train.json, qrecc_dev.json, qrecc_test.json.
OR-QUAC dataset: orquac_train.json, orquac_dev.json, orquac_test.json.
Format:

 {
    "Conversation_no": 3209,
    "Turn_no": 2,
    "Context": [
      "who is  finn m. w. caspersen?",
      "American financier and philanthropist."
    ],
    "Question": "where did he study?",
    "Gold_question": "",
    "Answer": "Peddie School, Brown University, and Harvard Law School.",
    "Page": "Finn M. W. Caspersen",
    "Section": "Early life and education",
    "Passage": {
      "id": "8114812",
      "text": "He later reflected that being Protestant was important. There was a kind of anti-Catholicism in the family. The family moved to homes in Andover, New Jersey, and Venice, Florida. Caspersen frequently visited Norway as a child, vacationing there during summers after 1947. Caspersen attended private schools until the ninth grade. He attended the Peddie School, a private preparatory school in Hightstown, New Jersey, and was graduated in 1959. Caspersen received a Bachelor of Arts (B.A.) degree from Brown University in 1963 and a law degree (LL.B.) from Harvard Law School in 1966.",
      "title": "Finn M. W. Caspersen [SEP] Early life and education"
    }
  }

3. Trie.

To implement the constrained generation in the LLM, we process all the corpus and store it in the trie structure.
You could use the scripts passage-level/generate_trie_dict.py and document-level/generate_trie_dict.py to obtain the trie for passages and documents, respectively.
You could also download our processed trie files.

trie_dict_t5-base_section_level.pkl is for the passage_level.  
trie_dict_t5-base_page_level.pkl is for the document_level.

4. Download.

You could download the above files via this link.

Model training

Passage_level

The script for training on the TOPIOCQA dataset is

    - python3 -m torch.distributed.launch --nproc_per_node 8 passage-level/train_query_encoder.py
      --do_train True
      --load_small False
      --fp16 False
      --num_train_epochs 40
      --per_gpu_train_batch_size 8
      --per_gpu_eval_batch_size 4
      --per_gpu_test_batch_size 2
      --overwrite_output_dir True
      --train_file $$AMLT_DATA_DIR/QA_pairs/topiocqa/topiocqa_train.json
      --dev_file $$AMLT_DATA_DIR/QA_pairs/topiocqa/topiocqa_dev.json
      --test_file $$AMLT_DATA_DIR/QA_pairs/topiocqa/topiocqa_test.json
      --corpus_path $$AMLT_DATA_DIR/full_wiki_segments.json
      --cache_dir $$AMLT_DATA_DIR/huggingface_cache/
      --trie_dict $$AMLT_DATA_DIR/trie_dict_t5-base_section_level.pkl
      --output_dir $$AMLT_OUTPUT_DIR/release_test/
      --learning_rate 1e-5
      --prepend_answers True
      --model_type t5-large
      --top_k 5
      --beam_size 5

The script for training on the QRECC dataset is

    - python3 -m torch.distributed.launch --nproc_per_node 8 passage-level/train_query_encoder.py
      --do_train True
      --load_small False
      --fp16 False
      --num_train_epochs 40
      --per_gpu_train_batch_size 8
      --per_gpu_eval_batch_size 4
      --per_gpu_test_batch_size 2
      --overwrite_output_dir True
      --train_file $$AMLT_DATA_DIR/QA_pairs/qrecc/qrecc_train.json
      --dev_file $$AMLT_DATA_DIR/QA_pairs/qrecc/qrecc_dev.json
      --test_file $$AMLT_DATA_DIR/QA_pairs/qrecc/qrecc_test.json
      --corpus_path $$AMLT_DATA_DIR/full_wiki_segments.json
      --cache_dir $$AMLT_DATA_DIR/huggingface_cache/
      --trie_dict $$AMLT_DATA_DIR/trie_dict_t5-base_section_level.pkl
      --output_dir $$AMLT_OUTPUT_DIR/release_test/
      --learning_rate 1e-5
      --prepend_answers True
      --model_type t5-large
      --top_k 5
      --beam_size 5

The script for training on the OR-QUAC dataset is

    - python3 -m torch.distributed.launch --nproc_per_node 8 passage-level/train_query_encoder.py
      --do_train True
      --load_small False
      --fp16 False
      --num_train_epochs 40
      --per_gpu_train_batch_size 8
      --per_gpu_eval_batch_size 4
      --per_gpu_test_batch_size 2
      --overwrite_output_dir True
      --train_file $$AMLT_DATA_DIR/QA_pairs/orquac/orquac_train.json
      --dev_file $$AMLT_DATA_DIR/QA_pairs/orquac/orquac_dev.json
      --test_file $$AMLT_DATA_DIR/QA_pairs/orquac/orquac_test.json
      --corpus_path $$AMLT_DATA_DIR/full_wiki_segments.json
      --cache_dir $$AMLT_DATA_DIR/huggingface_cache/
      --trie_dict $$AMLT_DATA_DIR/trie_dict_t5-base_section_level.pkl
      --output_dir $$AMLT_OUTPUT_DIR/release_test/
      --learning_rate 1e-5
      --prepend_answers False
      --model_type t5-large
      --top_k 5
      --beam_size 5

Document_level

The script for training on the TOPIOCQA dataset is

    - python3 -m torch.distributed.launch --nproc_per_node 8 document-level/train_query_encoder.py
      --do_train True
      --load_small False
      --fp16 False
      --num_train_epochs 40
      --per_gpu_train_batch_size 8
      --per_gpu_eval_batch_size 4
      --per_gpu_test_batch_size 2
      --overwrite_output_dir True
      --train_file $$AMLT_DATA_DIR/QA_pairs/topiocqa/topiocqa_train.json
      --dev_file $$AMLT_DATA_DIR/QA_pairs/topiocqa/topiocqa_dev.json
      --test_file $$AMLT_DATA_DIR/QA_pairs/topiocqa/topiocqa_test.json
      --corpus_path $$AMLT_DATA_DIR/full_wiki_document.json
      --cache_dir $$AMLT_DATA_DIR/huggingface_cache/
      --trie_dict $$AMLT_DATA_DIR/trie_dict_t5-base_page_level.pkl
      --output_dir $$AMLT_OUTPUT_DIR/release_test/
      --learning_rate 1e-5
      --prepend_answers True
      --model_type t5-large
      --top_k 5
      --beam_size 5

The script for training on the QRECC dataset is

    - python3 -m torch.distributed.launch --nproc_per_node 8 document-level/train_query_encoder.py
      --do_train True
      --load_small False
      --fp16 False
      --num_train_epochs 40
      --per_gpu_train_batch_size 8
      --per_gpu_eval_batch_size 4
      --per_gpu_test_batch_size 2
      --overwrite_output_dir True
      --train_file $$AMLT_DATA_DIR/QA_pairs/qrecc/qrecc_train.json
      --dev_file $$AMLT_DATA_DIR/QA_pairs/qrecc/qrecc_dev.json
      --test_file $$AMLT_DATA_DIR/QA_pairs/qrecc/qrecc_test.json
      --corpus_path $$AMLT_DATA_DIR/full_wiki_document.json
      --cache_dir $$AMLT_DATA_DIR/huggingface_cache/
      --trie_dict $$AMLT_DATA_DIR/trie_dict_t5-base_page_level.pkl
      --output_dir $$AMLT_OUTPUT_DIR/release_test/
      --learning_rate 1e-5
      --prepend_answers True
      --model_type t5-large
      --top_k 5
      --beam_size 5

The script for training on the OR-QUAC dataset is

    - python3 -m torch.distributed.launch --nproc_per_node 8 document-level/train_query_encoder.py
      --do_train True
      --load_small False
      --fp16 False
      --num_train_epochs 40
      --per_gpu_train_batch_size 8
      --per_gpu_eval_batch_size 4
      --per_gpu_test_batch_size 2
      --overwrite_output_dir True
      --train_file $$AMLT_DATA_DIR/QA_pairs/orquac/orquac_train.json
      --dev_file $$AMLT_DATA_DIR/QA_pairs/orquac/orquac_dev.json
      --test_file $$AMLT_DATA_DIR/QA_pairs/orquac/orquac_test.json
      --corpus_path $$AMLT_DATA_DIR/full_wiki_document.json
      --cache_dir $$AMLT_DATA_DIR/huggingface_cache/
      --trie_dict $$AMLT_DATA_DIR/trie_dict_t5-base_page_level.pkl
      --output_dir $$AMLT_OUTPUT_DIR/release_test/
      --learning_rate 1e-5
      --prepend_answers False
      --model_type t5-large
      --top_k 5
      --beam_size 5

We trained the models on 8*32GB NVIDIA V100 GPUs. We release our trained model checkpoints on the three datasets in this link.

Contact

If there is any problem, please email [email protected]. Please do not hesitate to email me directly as I do not frequently check GitHub issues.