Awesome

A Contrastive Learning Neural Model for Math Word Problems

This repository is the PyTorch implementation for the paper:

Seeking Patterns, Not just Memorizing Procedures: Contrastive Learning for Solving Math Word Problems

A Contrastive Learning Model

Requirements

• python 3
• PyTorch 1.8
• transformers 4.9.1

Usage

• Data download

  • Our processed data
    • Chinese dataset: Math_23K
    • English dataset: MathQA
    • Data-processing codes are also provided in tools/

• Install transformer library

  pip install transformers
  

• Download pretrained bert model and Emplace vocab_list

  • We use pretrained bert model from https://huggingface.co/

  • Add [NUM] to the original vocab_list of the pretrained bert model as vocab.txt in pretrained_models_vocab/ :

    [PAD]
    [num]
    [NUM]
    [unused3]
    [unused4]
    [unused5]
    [unused6]
    ...
    

• Directory structure

  Organize above-mentioned files like this.

  .
  ├── data/
  │   ├── MathQA_bert_token_test.json
  │   ├── ...
  │   └── pairs/
  │       ├── MathQA-MathQA-sample.json
  │       └── ...
  ├── pretrained_models/
  │   └── bert-base-chinese/
  │       ├── vocab.txt(emplaced)
  │       └── ...
  ├── src/
  ├── tools/
  ├── run_cl.py
  ├── run_ft.py
  ├──...
  
  

• Train

  We provide some train shell scripts, and give an example. If you want to write your own train script, please see the code for more details.

  • Train stage 1: contrastive learning

    train contrastive learning	mono-lingual-zh	mono-lingual-en	multi-lingual
    shell script	./train-cl-monolingual-zh.sh	./train-cl-monolingual-en.sh	./train-cl-multilingual.sh

  • Train stage 2: finetune

    finetune	mono-lingual-zh	mono-lingual-en	multi-lingual-zh	multi-lingual-en
    shell script	./train-ft-monolingual-zh.sh	./train-ft-monolingual-en.sh	./train_ft_multilingual-zh.sh	./train_ft_multilingual-en.sh

  • An example

    To train a multi-lingual contrastive learning model. You can first run this shell. The model will be saved to ./output

    bash train-cl-multilingual.sh
    

    Then you can finetune the above model in one specific language, using

    bash train_ft_multilingual-zh.sh
    

• Trained Model

  We also provide our Trained Model after stage 2.

  You can load a model by adding

  --model_reload_path model_monolingual_en 		# as an example
  

Results

Citation

@misc{li2021seeking,
      title={Seeking Patterns, Not just Memorizing Procedures: Contrastive Learning for Solving Math Word Problems}, 
      author={Zhongli Li and Wenxuan Zhang and Chao Yan and Qingyu Zhou and Chao Li and Hongzhi Liu and Yunbo Cao},
      year={2021},
      eprint={2110.08464},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}