Code of paper "AdaLoGN: Adaptive Logic Graph Network for Reasoning-Based Machine Reading Comprehension".

torch==1.7.1
dgl-cu101==0.6.1
stanza==1.2.3
transformers==4.5.0
networkx
nltk
scikit-learn
pylev

We use Graphene to extract EDUs. We put all the contexts and options line by line in a .txt file and follow the instructions of Graphene to get EDUs outputs. Or you can also use our preprocessed file under directory ReclorDataset/LogiDataset. We also provide cached file of preprocessed datas on Google Drive. Download and put them under directory ReclorDataset/LogiQADataset.

Checkpoints can be accessed on Google Drive.

export MODE=eval_only
bash scripts/LogiGraph_Roberta.sh /PATH/TO/RECLOR/CHECKPOINTS  ## ReClor evaluation
bash scripts/LogiGraph_Roberta_LogiQA.sh /PATH/TO/LOGIQA/CHECKPOINTS  ## LogiQA evaluation

For ReClor dataset, we submit prediction file on ReClor Leaderboard and AdaLoGN achieves Rank #10 on leaderboard (03/15/2022).

You can also install wandb and set export WANDB_DISABLED=false in training scripts to visualize the training process.

export MODE=do_train
bash scripts/LogiGraph_Roberta.sh /PATH/TO/ROBERTA/LARGE  ## ReClor training
bash scripts/LogiGraph_Roberta_LogiQA.sh /PATH/TO/ROBERTA/LARGE  ## LogiQA training

Please cite this paper kindly in your publications if it helps your research.

@inproceedings{li2022adalogn,
  title={AdaLoGN: Adaptive Logic Graph Network for Reasoning-Based Machine Reading Comprehension},
  author={Li, Xiao and Cheng, Gong and Chen, Ziheng and Sun, Yawei and Qu, Yuzhong},
  booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={7147--7161},
  year={2022}
}