APOLLO: An Optimized Training Approach for Long-form Numerical Reasoning (LREC-Coling 2024)

Our paper is accepted by LREC-Coling 2024 !🥳🥳🥳

This repo provides the code of APOLLO. In the paper, we adopt a number-aware negative sampling strategy in retriever to discriminate key numerical facts from others. Moreover, we design consistency-based reinforcement learning with target program augmentation, to increase program diversity and ultimately increase the execution accuracy.

---

How to cite

If you extend or use this work, please cite the paper where it was introduced:

@article{sun2022apollo,
  title={APOLLO: An Optimized Training Approach for Long-form Numerical Reasoning},
  author={Sun, Jiashuo and Zhang, Hang and Lin, Chen and Gong, Yeyun and Guo, Jian and Duan, Nan},
  journal={arXiv preprint arXiv:2212.07249},
  year={2022}
}

---

Leaderboard

The FinQA and ConvFinQA challenge leaderboard is on CodaLab https://codalab.lisn.upsaclay.fr/competitions/4138#results and https://codalab.lisn.upsaclay.fr/competitions/8582#results. APOLLO achieves rank 1 on both datasets.

---

Requirements

• pytorch 1.7.1
• huggingface transformers 4.4.2
• sympy 1.9

Main Results

Generator Results:

Retriever Results:

---

Code

Datasets

The FinQA and ConvFinQA datasets are in /APOLLO/dataset/.In FinQA, there are train,dev and test set. However, in ConvFinQA, there are only train and dev set.

The Retriever

Pre-process

Pre-processing is mainly to process the data of train, dev and test in advance, so that the training and testing stage can directly load these processed data. To pre-process these data, edit data_process.py to set your own project path and some hyperparameters. In particular, if you use number-aware negative sampling during training, edit and run number_aware.py to create the train_number_aware.json file in advance. Specifically, firstly run:

python number_aware.py

Secondly, run:

python data_process.py

Train

To train the retriever, you can run:

python -u -m torch.distributed.launch --nproc_per_node=2 --master_port=8889 Main.py\
--root_path "/APOLLO/" \
--model_save_name retriever-deberta \
--pretrained_model deberta \
--model_size microsoft/deberta-v3-large \
--mode train \
--neg_rate 3 \
--max_seq_length 512 --batch_size 8 --gradient_accumulation_steps 1 \
--learning_rate 2e-5 --epoch 50 --report 500 \
--features_dir /APOLLO/dataset/retriever/ \
--examples_dir /APOLLO/dataset/retriever/ \
--tags 1 \
--dataset_type finqa

Inference

To inference, you can run:

python -u -m torch.distributed.launch --nproc_per_node=1 --master_port=8899 Main.py\
--root_path "/APOLLO/" \
--model_save_name retriever-deberta \
--pretrained_model deberta \
--model_size microsoft/deberta-v3-large \
--mode inference \
--features_dir /APOLLO/dataset/retriever/ \
--examples_dir /APOLLO/dataset/retriever/ \
--saved_model_path "the path of your selected checkpoint in the training" \
--dataset_type finqa --tags 2 

Convert

To convert data for generator training, you can run:

python -u -m torch.distributed.launch --nproc_per_node=1 --master_port=8999 Main.py\
--root_path "/APOLLO/" \
--model_save_name retriever-deberta \
--pretrained_model deberta \
--model_size microsoft/deberta-v3-large \
--mode convert \
--dataset_type finqa --tags 3 

The Generator

Pre-process

To pre-process the data, you can edit and run:

python data_process.py

Train

To train the retriever, you can run:

python -u -m torch.distributed.launch --nproc_per_node=2 --master_port=7889 Main.py\
--root_path "/APOLLO/" \
--model_save_name generator-roberta-large \
--pretrained_model roberta \
--model_size roberta-large \
--mode train \
--retrieve_mode single --program_mode seq \
--max_seq_length 512 --batch_size 8 --gradient_accumulation_steps 1 \
--learning_rate 2e-5 --epoch 50 --max_program_length 30 --report 500 \
--features_dir /APOLLO/dataset/generator/ \
--examples_dir /APOLLO/dataset/generator/ \
--tags 1 \
--dataset_type finqa

Inference

To inference, you can run:

python -u -m torch.distributed.launch --nproc_per_node=1 --master_port=7899 Main.py\
--root_path "/APOLLO/" \
--model_save_name generator-roberta-large \
--pretrained_model roberta \
--model_size roberta-large \
--mode inference \
--features_dir /APOLLO/dataset/generator/ \
--examples_dir /APOLLO/dataset/generator/ \
--tags 2 \
--saved_model_path "the path of your selected checkpoint in the training" \
--dataset_type finqa

Consistency-based Reinforcement learning

In order to adopt reinforcement learning, you need to train a model in the above process. Then, you can run:

python -u -m torch.distributed.launch --nproc_per_node=2 --master_port=6899 Main.py\
--root_path "/APOLLO/" \
--model_save_name generator-roberta-large \
--pretrained_model roberta \
--model_size roberta-large \
--mode train \
--features_dir /APOLLO/dataset/generator/ \
--examples_dir /APOLLO/dataset/generator/ \
--tags 3 \
--saved_model_path "the path of your selected checkpoint in the supervised training" \
--dataset_type finqa --rl

The test of reinforcement learning is the same as the inference of supervised training.

Target Program Augmentation

In order to adopt target program augmentation, you need to cd in /TPA Then, you can run:

python TPA_Switch.py
python TPA_Add_Subtract.py
python TPA_Multiply_Divide.py
python TPA_Mul-Div.py

to create TPA dataset. Then, you need to edit and run data_process.py in /Generator to create features:

python data_process.py

In particular, you need to edit:

examples_dir = "/APOLLO/dataset/generator/"
features_dir = "/APOLLO/dataset/generator/"
if dataset_type == "finqa":
    train_file = root_path + "dataset/FinQA/train_retrieve_output.json"
else:
    train_file = root_path + "dataset/ConvFinQA/train_retrieve_output.json"
f = open(os.path.join(examples_dir, 'train_examples.pickle'), 'wb')
f = open(os.path.join(examples_dir, 'dev_examples.pickle'), 'wb')
f = open(os.path.join(features_dir, 'train_features.pickle'), 'wb')
f = open(os.path.join(features_dir, 'dev_features.pickle'), 'wb')
f = open(os.path.join(examples_dir, 'test_examples.pickle'), 'wb')
f = open(os.path.join(features_dir, 'test_features.pickle'), 'wb')

to

examples_dir = "/APOLLO/dataset/generator_tpa/"
features_dir = "/APOLLO/dataset/generator_tpa/"
if dataset_type == "finqa":
    train_file = root_path + "dataset/FinQA/train_TPA_Switch.json"  # or other TPA methods
else:
    train_file = root_path + "dataset/ConvFinQA/train_TPA_Switch.json"  # or other TPA methods
f = open(os.path.join(examples_dir, 'train_examples_switch.pickle'), 'wb')
f = open(os.path.join(examples_dir, 'dev_examples_switch.pickle'), 'wb')
f = open(os.path.join(features_dir, 'train_features_switch.pickle'), 'wb')
f = open(os.path.join(features_dir, 'dev_features_switch.pickle'), 'wb')
f = open(os.path.join(examples_dir, 'test_examples_switch.pickle'), 'wb')
f = open(os.path.join(features_dir, 'test_features_switch.pickle'), 'wb')

Then, you can run this command in /Generator:

python -u -m torch.distributed.launch --nproc_per_node=2 --master_port=5899 Main.py\
--root_path "/APOLLO/" \
--model_save_name generator-roberta-large \
--pretrained_model roberta \
--model_size roberta-large \
--mode train \
--features_dir /APOLLO/dataset/generator_tpa/ \
--examples_dir /APOLLO/dataset/generator_tpa/ \
--tags 4 \
--saved_model_path "the path of your selected checkpoint in the supervised training" \
--dataset_type finqa --tpa \
--tpa_methods switch

The test of TPA is the same as the inference of supervised training.