amy-hyunji / generative-multihop-retrieval

Generative-Multihop-Retrieval

This is the official github repository for Generative Multihop Retrieval, accepted at EMNLP 2022.

Use the following to cite our paper:

@inproceedings{Lee2022GenerativeMR,
  title={Generative Multi-hop Retrieval},
  author={Hyunji Lee and Sohee Yang and Hanseok Oh and Minjoon Seo},
  booktitle={Conference on Empirical Methods in Natural Language Processing},
  year={2022}
}

In order to reproduce our results, take the following steps:

1. Create conda environment and install requirements

conda create -n gmr python=3.7 && conda activate gmr
pip install -r requirements.txt

2. Download the data used for the experiments.

1. EntailmentBank
2. StrategyQA
3. Explagraphs-Open
4. RuleTaker-Open
5. HotpotQA

Detailed preprocessing code is in process

3. Set Config

Components in each configurations file

{
    "output_dir" (str): output directory to save all ckpt
    "dataset" (str): path to the dataset
    "model" (str): ['t5-base', 't5-large']
    "max_input_length" (int): max input length
    "max_output_length" (int): max output length
    "learning_rate" (float): learning rate (default: 1e-4)
    "lr_scheduler" (str): scheduler ['constant', 'exponential']
    "num_train_epochs" (int): training epoch (default: 150 with early stopping)
    "train_batch_size" (int): batch size for eval (default: 120)
    "eval_batch_size" (int): batch size for eval
    "gradient_accumulation_steps" (int): gradient accumulation size
    "n_gpu" (int): # of gpus
    "num_workers" (int): # of cpus
    "resume_from_checkpoint" (str): when resume training from ckpt
    "seed" (int): seed (default: 101)
    "check_val_every_n_epoch" (int): freq of validation (default: 3)
    "train_file" (str): path to train dataset
    "dev_file" (str): path to dev dataset
    "test_file" (str): path to test dataset
    "prefix_tree" (str): path to prefix tree
    "constrained_decoding" (bool): [true (default) | false]
    "do_train" (bool): [true | false], true if in train step else false
    "do_test" (bool): [true | false], true if in test step else false
    "setting" (str): ["ret_fixed", "ret_dynamic", "LM_mem", "multihop_mem"]
    "test_model_path" (str): path to test ckpt
    "beam_size" (int): beam size
    "ret_num" (int): # of sequences to return
    "wandb_log" (bool): log in wandb
    "wandb_project" (str): project name in wandb
    "wandb_run_name" (str): run name in wandb
    "accelerator" (str): [ddp (default) | deepspeed] 
}

4. Train Model

• config setting:

  do_train: true
  do_test: false
  

Memorization Step

python train.py --config config/*

We are in the process of releasing code for memorization step!

• Language Model Memorization
  • setting in config should be LM_mem
• Multihop Memorization
  • setting in config should be multihop_mem

Retrieval Step

python train.py --config config/*

• Fixed Setting
  • setting in config should be ret_fixed
• Dynamic Setting
  • setting in config should be ret_dynamic

5. Test Model

• config setting:

  do_train: true
  do_test: false
  

• run

  python train.py --config config/*
  

generative-multihop-retrieval's Issues

Hello, I'm interested in your work. When will the code be made public?

dataset

Hello, what form is the dataset structured in, your readme doesn't make it clear

Will the code be released?

and checkpoint