MERIt: Meta-Path Guided Contrastive Learning for Logical Reasoning

This is the pytorch implementation of the paper:

MERIt: Meta-Path Guided Contrastive Learning for Logical Reasoning. Fangkai Jiao, Yangyang Guo, Xuemeng Song, Liqiang Nie. Findings of ACL. 2022.

paper link

Project Structure

|--- conf/   // The configs of all experiments in yaml.
|--- dataset/   // The classes or functions to convert raw text inputs as tensor and utils for batch sampling.
|--- experiments/   // We may provide the prediction results on the datasets, e.g., the .npy file that can be submitted to ReClor leaderboard.
|--- general_util/  // Metrics and training utils.
|--- models/    // The transformers for pre-training and fine-tuning.
|--- modules/
|--- preprocess/    // The scripts to pre-process Wikipedia documents for pre-training.
|--- scripts/   // The bash scripts of our experiments.
|--- reclor_trainer....py   // Trainers for fine-tuning.
|--- trainer_base....py     // Trainers for pre-training.

Requirements

For general requirements, please follow the requirements.txt file.

Some special notes:

The fairscale is used for fast and memory efficient distributed training, and it's not necessary.
NVIDIA Apex is required if you want to use the FusedLAMB optimizer for pre-training. It's also not necessary since you can use AdamW.
RoBERTa-large requires at least 12GB memory on single GPU, ALBERT-xxlarge requires at least 14GB, and we use A100 GPU for Deberta-v2-xlarge pre-training. Using fairscale can help reduce the memory requirements with more GPUs. We have tried to use Deepspeed for Deberta pre-training through CPU-offload but failed.

Pre-training

The pre-trained models are available at HuggingFace:

For all model checkpoints (including models of different pre-training steps and the models for ablation studies), we will provide them later.

You can also pre-train the models by yourself as following.

Data Preprocess

Our pre-training procedure uses the data pre-processed by Qin et al., which includes the entities and the distantly annotated relations in Wikipedia. You can also download it from here.

To further pre-process the data, run:

python preprocess/wiki_entity_path_preprocess_v7.py --input_file <glob path for input data> --output_dir <output path>

The processed data can also be downloaded from here.

Running

We use hydra to manage the configuration of our experiments. The configs for pre-training are listed below:

# RoBERTa-large-v1
conf/wiki_path_large_v8_2_aug1.yaml

# RoBERTa-large-v2
conf/roberta/wiki_path_large_v8_2_aug1_fix_v3.yaml

# ALBERT-v2-xxlarge
conf/wiki_path_albert_v8_2_aug1.yaml

# DeBERTa-v2-xlarge
conf/deberta_v2/wiki_path_deberta_v8_2_aug1.yaml

# DeBERTa-v2-xxlarge
conf/deberta_v2/wiki_path_deberta_v8_2_aug1_xxlarge.yaml

To run pre-training:

python -m torch.distributed.launch --nproc_per_node N trainer_base_mul_v2.py -cp <directory of config> -cn <name of the config file>

DeepSpeed is also supported by using trainer_base_mul_ds_v1.py and relevant config is specified in the ds_config field of the config.

Fine-tuning

To run fine-tuning:

python -m torch.distributed.launch --nproc_per_node N reclor_trainer_base_v2.py -cp <directory of config> -cn <name of the config file>

The configs for different experiments are listed as following (the details of our used GPUs and the seed to achieve the best accuracy on the test set are also provided):

MERIt-roberta-large-v1

ReClor
- config: conf/ft_v8_2_2_1aug_ctx_1k.yaml (4 * 2080Ti, seed=4321)

LogiQA
- config: conf/logiqa/f_logiqa_large_v8_2_1aug_ctx.yaml (1 * Tesla T4, seed=44)

MERIt-roberta-large-v1 + Prompt

ReClor
- config: conf/p_ft_v8_2_2_1aug_ctx_1k.yaml (2 * 2080Ti, seed=43)

LogiQA
- config: conf/logiqa/pf_logiqa_large_v8_2_1aug_ctx.yaml (1 * Tesla T4, seed=44)

MERIt-roberta-large-v2

ReClor
- config: scripts/steps/run_steps_fix.sh (2 * 2080Ti, seed=43)

MERIt-roberta-large-v2 + Prompt

ReClor
- config: conf/roberta/p_ft_v8_2_2_1aug_ctx_fix_v3.yaml (2 * 2080Ti, seed=4321)

MERIt-alberta-v2-xxlarge

ReClor
- config: conf/albert/albert_ft_v822_1aug_ctx (2 * TitanXP, seed=4321)

LogiQA
- config: conf/albert/logiqa_albert_ft_v822_1aug_ctx.yaml (1 * 2080Ti, seed=42)

MERIt-alberta-v2-xxlarge + Prompt

ReClor
- config: conf/albert/albert_pt_v822_1aug_ctx.yaml (2 * TeslaT4, seed=44)

LogiQA
- config: conf/albert/logiqa_albert_pt_v822_1aug_ctx.yaml (1 * 2080Ti, seed=43)

MERIt-deberta-v2-xlarge

ReClor
- config: conf/deberta_v2/deberta_ft_path_v1_4_2.yaml (2 * A100, seed=42)

MERIt-deberta-v2-xxlarge

ReClor
- config: conf/deberta_v2/deberta_xxlarge_ft_path.yaml (2 * A100, seed=42)

Citation

If the paper and code are helpful, please kindly cite our paper:

@inproceedings{Jiao22merit,
  author    = {Fangkai Jiao and
               Yangyang Guo and
               Xuemeng Song and
               Liqiang Nie},
  title     = {{MERI}t: Meta-Path Guided Contrastive Learning for Logical Reasoning},
  booktitle = {Findings of ACL},
  publisher = {{ACL}},
  year      = {2022}
}

SparkJiao/MERIt