COMUS

This is the official PyTorch implementation for the paper:

Zheng Gong*, Kun Zhou*, Xin Zhao, Jing Sha, Shijin Wang, Ji-Rong Wen. Continual Pre-training of Language Models for Math Problem Understanding with Syntax-Aware Memory Network. ACL 2022.

Overview

We propose COMUS, a new approach to continually pre-train language models for math problem understanding with syntax-aware memory network. We construct math syntax graph to model the structural semantic information, and then design the syntax-aware memory networks to deeply fuse the features from the graph and text. We finally devise three continual pre-training tasks to further align and fuse the representations of the text and math syntax graph.

Requirements

# for model
torch==1.10.0
transformers==4.6.0
datasets==1.1.3
dgl==0.8.x
# for data process
jieba
sympy
bs4
stanza
timeout_decorator

Dataset

Datasets cannot be shared temporarily for some commercial reasons. We put the data pre-processing code in the data folder for reference. In general, the pre-processing of the data consists of the following parts:

• Getting a clean corpus of math problems with formula location identifiers (like $ a + b $).
• Parsing the formulas into operator trees. In this project, we made some modifications to TangentS to accomplish this step (using sympy instead of latexml to perform latex-to-mathml formatting). We suggest referring to the original project's code (TangentS) for this step.
• Dependency parsing for math problem texts. We use stanza to do this.
• Combining the dependency parsing tree and the operator tree to obtain the math syntax graph.
• Organizing the data (including text and math syntax graph) into Dataset format.

We provide two fake samples in data/example_data.json. You can process them with following codes and see the data generated in each stage.

cd data && python preprocess.py

Note: You need to modify some codes if your data are not in Chinese. Check line 20 of data/data_utils.py and line 31~40 of data/preprocess.py

Training

Base Model

Please download the initial model from https://huggingface.co/models.

• Make sure your initial model is bert-like (bert or roberta). Otherwise you need to modify the training and model code.
• We run this project with bert-base. You can adjust the hyperparameters of GAT in model/config.py if you want to try larger model.

Scripts

You can run pre-training with single GPU by:

bash scripts/run_pretrain.sh

or run distributed data paralle pre-training with multiple GPUs by:

bash scripts/run_pretrain_ddp.sh

Arguments

You can check more details about training arguments in the official docs of huggingface. We explain some special arguments here.

• model_name_or_path - Directory of model checkpoint for weights initialization. Put your downloaded base model here.
• data_path - Your pre-processed training data saved in Dataset format (line 103 of data/preprocess.py).
• add_token_path - There may be some important words in your corpus that cannot be correctly split by the tokenizer of the pre-trained model, such as mathematical symbols. You can add them to the vocab by this argument and train the embedding from scratch.
• graph_vocab_path - The node set of math syntax graph.

Note: You can choose any file type to store graph vocab and additional tokens. Just check and modify the loading code in pretrain.py.

Citation

Please consider citing our paper if you use our codes.

@inproceedings{gong-etal-2022-continual,
    title = "Continual Pre-training of Language Models for Math Problem Understanding with Syntax-Aware Memory Network",
    author = "Gong, Zheng  and
      Zhou, Kun  and
      Zhao, Xin  and
      Sha, Jing  and
      Wang, Shijin  and
      Wen, Ji-Rong",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.408",
    doi = "10.18653/v1/2022.acl-long.408",
    pages = "5923--5933",
}