README.md

MathGPT

Code for the paper Tree-Based Representation and Generation of Natural and Mathematical Language

If you end up using this code in your research, please cite us like:

@inproceedings{scarlatos-lan-2023-tree,
    title = "Tree-Based Representation and Generation of Natural and Mathematical Language",
    author = "Scarlatos, Alexander and Lan, Andrew",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.205",
    pages = "3714--3730",
}

Setup

Python Environment

Ensure Python3 is installed (this code was tested on v3.9.1).

Create virtual environment

python3 -m venv <env_name>
source <env_name>/bin/activate

Install libraries

python3 -m pip install -r requirements.txt

Make TangentCFT available in Python path

export PYTHONPATH=..:../TangentCFT/

External Dependencies

The following need to be installed for full functionality. LaTeXML is only required to run pre-processing.

TangentCFT

Download TangentCFT to the folder above the root of this repo: https://github.com/BehroozMansouri/TangentCFT/tree/2b189dff67d6d3b15e323358921bdbca779bfcd9

Note that some small fixes were made to TangentCFT, so the file semantic_symbol.py is copied from there with some changes and is overloaded automatically.

LaTeXML

https://math.nist.gov/~BMiller/LaTeXML/get.html

NLG-Eval

https://github.com/Maluuba/nlg-eval

Known installation issue: Maluuba/nlg-eval#61

Data

Here are links to the datasets required for the following tasks. For each, ensure the dataset's root folder is above the root of this repo.

• Pre-Training: https://ntcir-math.nii.ac.jp/data/
  • just the Wikipedia Corpus
• Headline Generation: https://github.com/yuankepku/MathSum
• Equation Extraction: https://ai.tencent.com/ailab/nlp/dialogue/#datasets
  • The dataset will have to be translated to English; Google Translate is acceptable.
  • See pre_process.py->process_mwp_data for desired final data format.
• Student Action Prediction: https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=660
• GSM8K: https://github.com/openai/grade-school-math
• MATH: https://github.com/hendrycks/math

The datasets for the following tasks cannot be released publicly:

• Answer Scoring
• Feedback Generation

Run

The starting point for all code is __main__.py, and you can see a list of command line options by running:

python3 __main__.py --help

Default values can be found in the TrainOptions constructor in utils.py.

Here is the typical workflow to replicate our experiments:

• Pre-process the Wikipedia dataset (this step also constructs the vocabulary, which is needed for all following steps)
• Pre-train a MathGPT model
• Pre-process the downstream dataset
• Run cross-validation on the downstream dataset, which for each fold:
  • Fine-tunes the pre-trained MathGPT model on the downstream dataset
  • Runs evaluation on the downstream dataset's test set