This is the official repo for the paper "A Fusion Network of Sequence-to-Sequence Model and Tree-based Decoder for Math Word Problems".
The checkpoints of MWP-BERT we used can be obtained from https://drive.google.com/drive/folders/1staS6zCiAarNNMz9GhckNGkc_OD2HL1O?usp=drive_link
Liang, Zhengwen et al., MWP-BERT: Numeracy-Augmented Pre-training for Math Word Problem Solving, NAACL 2022
Run the following to train the model and reproduce the results of Train-Test:
python -u run_bert2bert_tree_attributes.py
Run the following to train the model and reproduce the results of 5-Fold Cross-Validation:
python -u cross_valid_bert2bert_tree.py
Run the following to train the model and reproduce the results of Train-Test:
python -u run_wape.py
We reproduce the main results of BERT2BERT-Tree in the following tables:
| Dataset | Euqation accuracy | Value accuracy |
|---|---|---|
| Math23K (train/test) | 73.2 | 85.9 |
| Math23K (5-fold cross-validation) | 72.7 | 84.7 |
| Math23K (train with Ape-clean) | 77.7 | 91.5 |
| Ape-clean (train/test) | 68.0 | 81.8 |
The data used by the bert2bert model is uploaded to data/math23k/.
Wang, Yan, et al., Deep Neural Solver for Math Word Problems (Math23K first introduced), EMNLP 2017
The Math23K and Ape-clean data for bert2bert-tree is in data/.
Moreover, the elements of Ape-clean are sampled from Ape-210K as listed in (data/ape_simple_id.txt) and (data/ape_simple_test_id.txt)
Zhao, Wei, et al., Ape210K: A Large-Scale and Template-Rich Dataset of Math Word Problems, arXiv:2009.11506
Build the entity attributes for your data:
The process relies on Stanford CoreNLP, follow this guide to install the PyPI package and download the Chinese model package.
-
Follow the instruction Recut the word in
attributes_building.ipynbto prepare the data of entity attributes generating. -
Execute the Extract entity attributes of numbers in
attributes_building.ipynbto generate the corresponding number entity attributes.
If you find this work useful, please cite our paper:
@article{}