ReClor: A Reading Comprehension Dataset Requiring Logical Reasoning (ICLR 2020)

This repository contains PyTorch code for the paper: Weihao Yu*, Zihang Jiang*, Yanfei Dong, and Jiashi Feng, ReClor: A Reading Comprehension Dataset Requiring Logical Reasoning, ICLR 2020 (* equal contribution). For the data and more information, please check out the project page.

Setting up and using the repo

1. Set up the environment. Install Python3.5+, PyTorch 1.0+, Transformers and apex. I recommend the Anaconda distribution to set up Python environment. Refer to pytorch.org to install PyTorch. Then install Transformers package by

pip install transformers==2.3.0

Then refer to apex to install Nvidia apex for mixed precision training.

2. Clone the repo by

git clone https://github.com/yuweihao/reclor.git

3. Get the dataset. Download the dataset from this Download Link. The unzip password is for_non-commercial_research_purpose_only. Or you can use the following command to unzip the file:

mkdir reclor_data && unzip -P for_non-commercial_research_purpose_only -d reclor_data reclor_data.zip

4. Run the scripts in the main directory by such as

sh scripts/run_roberta_large.sh

5. (Optional) After running the script, you can find best_dev_results.txt in the checkpoint directory which record the best result on validation set, and the predicted file test_preds.npy for testing set that you can submit to the EvalAI leaderboard to see the accuracy. The predicted file test_preds.npy contains [3, 1, 0, 2, ...] where i-th element represents the predicted label of i-th question with "id_string": "test_i" in test.json.

Results

We obtain the following results with Ubuntu 16.04, NVIDIA driver 430, PyTorch 1.3.1, cudatoolkit 10.1, numpy 1.17.4, NVIDIA apex, and NVIDIA TITAN RTX GPU (we find the results are different between TITAN RTX and GeForce RTX 2080TI when runing large models).

Model	Val	Test	Test-E	Test-H
bert-base	54.6	47.3	71.6	28.2
bert-large	53.8	49.8	72.0	32.3
xlnet-base	55.8	50.4	75.2	32.9
xlnet-large	62.0	56.0	75.7	40.5
roberta-base	55.0	48.5	71.1	30.7
roberta-large	62.6	55.6	75.5	40.0

If you could not obtain similar performance in your environment and device, maybe you can try different random seeds.

Bibtex

@inproceedings{yu2020reclor,
        author = {Yu, Weihao and Jiang, Zihang and Dong, Yanfei and Feng, Jiashi},
        title = {ReClor: A Reading Comprehension Dataset Requiring Logical Reasoning},
        booktitle = {International Conference on Learning Representations (ICLR)},
        month = {April},
        year = {2020}
}

Acknowledgment

Weihao Yu would like to thank TPU Research Cloud (TRC) program for the support of partial computational resources.