ReadMe

This is the source code for the paper Weakly Supervised Formula Learner for Solving Mathematical Problems.

Environment

The code was written and verified on Python 3.8 and PyTorch 1.6.0.

The other required libraries include :

• json
• numpy

Getting start

Data preparation

We provide the program regularize.py for data pre-processing.

Math23K

For Math23K, download Math_23K.json from https://github.com/ShichaoSun/math_seq2tree and put it into ./data/math23/. Then run :

python ./data/math23/regularize.py

MathQA

For MathQA, download MathQA.zip from https://math-qa.github.io/. Extract train.json and test.json to ./data/mathqa/. Then run :

python ./data/mathqa/regularize.py

Training

To start the training, run :

python ./train.py

The whole learning process may take a long time. You can interrupt it at any time you like by a keyboard interruption, then the search log and model weights will be saved to ./saves/$dataset$/. The learning can be restarted by assigning the load_search_log and policynet_load_weight in config.py.

Testing

To test the model, assign the policynet_load_weight in config.py. Then run :

python ./test.py

Using pre-trained word vectors

We provide the option to load the pre-trained word vectors. To activate this option, switch policynet_load_word_embedding to True in config.py.

Math23K

For Math23K, download the 300d word vectors from https://github.com/Embedding/Chinese-Word-Vectors. Extract the raw .txt file to ./data/math23/. Then run :

python ./data/math23/generate_weight.py

MathQA

For MathQA, download the 300d word vectors from https://nlp.stanford.edu/projects/glove/. Extract the raw .txt file to ./data/mathqa/. Then run :

python ./data/mathqa/generate_weight.py