An Abductive Learning Framework For Automatically Solving Math Problems

This code is our implementation of paper An Abductive Learning Framework For Automatically Solving Math Problems, which is based on the famous NLP framework AllenNLP.

Citation: "Yangyang Hu and Yang Yu. Enhancing neural mathematical reasoning by abductive combination with symbolic library. In: ICML 2020 Workshop on Bridge Between Perception and Reasoning: Graph Neural Networks & Beyond, Vienna, Austria, 2020". https://arxiv.org/abs/2203.14487

Video presentation: https://agit.ai/Polixir/ABL-sym

Online demonstration: http://math.polixir.ai/

install

pip install -r requirements.txt

dataset format

We experiment on the Mathematics Dataset. The data set is preprocessed into the following form, where one line is the description of the problem, the next line is separated by ###, the first half is the answer, the second half is the program(may not have).

What is the first derivative of 5018w**4 + 15w3 + 15w + 12680810 wrt w?
20072w3 + 45*w**2 + 15###pos#3 pos#8 pos#6 argc#3 api#diff extra#@end@
Calculate the remainder when 7687189 is divided by 3441.
3436###pos#8 pos#4 argc#2 api#Mod extra#@end@

models

We release 3 models(extraction code: eg39)

• program-searching
• abl-sym
• abl-sym-tp

search

To execute program-searching procedure, use warm_up_search.py, program_search.py, curriculum_search.py. During the execution process, the searched programs need to be used to train the program-searching model, and the programs generated by the search model will further help the search process. For the training process, see the training section below.

python dmmath/search/warm_up_search.py \
    --data ...
python dmmath/search/program_search.py \
    --data ...
python dmmath/search/curriculum_search.py \
    --model .. \
    --data ...

training

To train ABL-sym model, use train_dist_transformer.py, the workdir directory should include experiment.config, which is used for training configuration.

python dmmath/nlp/train_dist_transformer.py \
    ---workdir ... \
    --local_rank ...

prediction

To predict results, use predict_mix_transformer.py, the data should be preprocessed as above.

python dmmath/nlp/predict_mix_transformer.py \
    --model ... \
    --data ... \
    --result_dir ... \
    --target_limit_decode_steps ... \
    --batch_size ... \
    --cuda ...

evaluation

To evaluate ABL-Sym on Mathematics Dataset, use evaluate_mix_transformer.py, the result_path is the prediction result path using predict_mix_transformer.py:

python dmmath/nlp/evaluate_mix_transformer.py \
    --result ...