Code for EMNLP 2021 Paper "Recall and Learn: A Memory-augmented Solver for Math Word Problems".

environment

python3.6, pytorch1.2
You can install related packages directly through "pip install requirements.txt" or
create virtual environment with required package by "conda env create -f REAL.yaml".

preprocess data

python3 memory_module.py

train:

 python3 run.py --is_train --train_batch_size 6 --num_train_epochs 80 \
    --start_lr_decay_epoch 40 --dataset math23k --topk 1 \
    --add_copynet --add_memory_module --is_equ_norm 

test:

python3 run.py --eval_batch_size 6  --dataset math23k --topk 1 \
    --add_copynet --add_memory_module --is_equ_norm 

parallel

The project supports multi GPU parallel by setting CUDA_VISIBLE_DEVICES.

result

We don't spend more energy on adjusting the parameters of the model. You can get better results through parameter adjustment.

top1 result:

<img width="500" height="300" src="https://github.com/sfeng-m/REAL4MWP/blob/master/images/top1_result.png" />

topk result:

Although REAL is trained with only a retrieved question, we still have the flexibility to adjust the number of retrieved questions at the testing stage by modifying the value of topk, which can affect the model’s performance.

<img width="500" height="300" src="https://github.com/sfeng-m/REAL4MWP/blob/master/images/topk_result.png" />

In addition, It's simple to train topk retrieved questions by modifying the value of topk at training stage, which can obtain better result, though it is not show in our paper.

Improve

Further, we improve the effectiveness of REAL model to solve math work problems(MWP) by optimizing the memory module.
More details please see our NIPS2021 Paper on MATHAI4ED Workshop: "REAL2: An End-to-end Memory-augmented Solver for Math Word Problems".

Citation

If the paper or the code helps you, please cite the paper in the following format :

@inproceedings{huang-etal-2021-recall-learn,
    title = "Recall and Learn: A Memory-augmented Solver for Math Word Problems",
    author = "Huang, Shifeng and Wang, Jiawei and Xu, Jiao and Cao, Da and Yang, Ming",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    year = "2021"
    }

Acknowledgments

Our code is based on unilm . We thank the authors for their wonderful open-source efforts. We use the same license as unilm.