Learning by Analogy: Diverse Questions Generation in Math Word Problem

Overview

Diverse questions generation framework (DQGF) can automatically generate diverse questions with their corre- 141 sponding equations for a given MWP.

Data

• DiverseMath23k: data/DiverseMath23k consists of our generated data on Math23k. Specifically, DiverseMath23k-train/dev/text.json is the dataset we reported in paper. Furthermore, we give the DiverseMath23k-original.json, which is a dataset generated by DQGF but not filtered. It has 28w diverse MWP data, you can use a strong filter (like ChatGPT) to get clean them.

• DiverseMath23k: data/Unbiased_DQGF consists of our generated data on Unbiased-source.

Equations Generator

#process data
python turnNum2Sym.py
#generate equations
python get_generate.py
#process equation file to the format in Question Generator
python process.py

Question Generater

Filter

• In this paper, we use MWP-bert as our filter.

Citation

@inproceedings{zhou-etal-2023-learning-analogy,
    title = "Learning by Analogy: Diverse Questions Generation in Math Word Problem",
    author = "Zhou, Zihao  and Ning, Maizhen  and Wang, Qiufeng  and Yao, Jie  and Wang, Wei  and Huang, Xiaowei  and Huang, Kaizhu",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
    year = "2023",
    pages = "11091--11104",
}