shichaosun / math_seq2tree Goto Github PK

View Code? Open in Web Editor NEW

141.0 5.0 38.0 3.16 MB

Seq2Tree model for Solving Math Word Problems

License: GNU General Public License v3.0

Python 100.00%

seq2tree math23k recursive-neural-network math

math_seq2tree's Introduction

A Goal-Driven Tree-Structured Neural Model for Math Word Problems

This repository is the PyTorch implementation for the IJCAI 2019 accepted paper:

Zhipeng Xie* and Shichao Sun*, A Goal-Driven Tree-Structured Neural Model for Math Word Problems IJCAI 2019.

* indicates equal contribution.

Seq2Tree Model

A Seq2Tree Neural Network containing top-down Recursive Neural Network and bottom-up Recursive Neural Network

Requirements

python 3
PyTorch 0.4.1

Train and Test

Math23K:

python3 run_seq2tree.py

Results

Model	Accuracy
Hybrid model w/ SNI	64.7%
Ensemble model w/ EN	68.4%
Seq2Tree w/o Bottom-up RvNN	70.0%
Seq2Tree	74.3%

Citation

@inproceedings{ijcai2019-736,
  title     = {A Goal-Driven Tree-Structured Neural Model for Math Word Problems},
  author    = {Xie, Zhipeng and Sun, Shichao},
  booktitle = {Proceedings of the Twenty-Eighth International Joint Conference on
               Artificial Intelligence, {IJCAI-19}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  pages     = {5299--5305},
  year      = {2019},
  month     = {7},
  doi       = {10.24963/ijcai.2019/736},
  url       = {https://doi.org/10.24963/ijcai.2019/736},
}

math_seq2tree's People

Contributors

Stargazers

Watchers

math_seq2tree's Issues

Train and test on the AllArith and MAWPS datasets

Hi, It looks like you released codes to train and test the Math23K dataset only. How can I train and test on the AllArith dataset (and MAWPS)?

源代码细节

作者您好，我在阅读您的代码时不知道num_stack的作用是什么。为什么式子里面有的数字没有被替换成N*呢？
希望能解答下，感谢：）

` for pair in pairs_trained:
num_stack = []
for word in pair[1]:
temp_num = []
flag_not = True
if word not in output_lang.index2word:
flag_not = False
for i, j in enumerate(pair[2]):
if j == word:
temp_num.append(i)

        if not flag_not and len(temp_num) != 0:
            num_stack.append(temp_num)
        if not flag_not and len(temp_num) == 0:
            num_stack.append([_ for _ in range(len(pair[2]))])

    num_stack.reverse()
    input_cell = indexes_from_sentence(input_lang, pair[0])
    output_cell = indexes_from_sentence(output_lang, pair[1], tree)`

Training time of Math23k for an epoch?

Hi,

May I ask how long did you train Math23K for an epoch? Your paper did not mention the concrete training time.

I tried to reimplement this paper. My dataset (4936 instances) needs around 10 min to run an epoch (I modified the data loader as it is kind of outdated), does that make sense?

Thank you

The code doesn't work

Hello GTS author, I just started running the source code and got an error message:
(UserWarning: masked_ fill_ received a mask with dtype torch. uint8, this behavior is now deprecated, please use a mask with dtype torch. bool instead.)
But when I replace it, the results will only be displayed to fold: 1, epoch: 1. There is no result after that. Find the solution, Thank you!