Multilingual-Deep-Neural-Math-Word-Problem-Solver

In this project, we develop a language agnostic math word problem solver using deep learning-based methods. Subsets of two multilingual large-scale datasets, Math23K (Chinese language) and Dolphin-S (English language) are used to train and test the language-agnostic model. Detailed descriptions can be found here.

Data is prepared in the following way:
The notebooks mentioned here are in Miscellaneous folder.

• Dolphin subset is prepared using Initial_Data_Cleaning.ipynb
• Dolphin is preprocessed in Preprocess_Dolphin.ipynb
• Math23K is preprocessed in Preprocess_Math23K.ipynb
• Dolphin is then replicated using SecondStage/replicate.py.
• LAMP32K is prepared in LAMP32K.ipynb
• LAMP32K is split in train,validation,test and postfix template preparation is in Split_Postfix.ipynb

All generated files including the dataset and splits are provided in the data/ folder.

Generator:
The code in FirstStage/src contains the model that will generate equations. It can be run using the command

python main.py --cuda-use --checkpoint-dir-name params_12 --mode 0 --teacher-forcing-ratio 0.5 --input-dropout 0.4 --encoder-hidden-size 512 --decoder-hidden-size 1024 --generator 1

This will create 3 files with generated equations for train, validation and test datasets. Sample files are provided in Results/GeneratorModel.

Predictor:
The code in SecondStage takes template equations and predicts operators between them. In order to train the model, set predict flag to False. To run the trained Predictor model, set the flag to True. To predict, downlaod the trained model from here to SecondStage/data/

python main.py true #predict
python main.py false #train

This will also generate sample files with predictions along with the ground truth.

GenPred:
In this, we generate equation templates as opposed to entire equation in Generator. The command to train this model is

python main.py --cuda-use --checkpoint-dir-name params_12 --mode 0 --teacher-forcing-ratio 0.5 --input-dropout 0.4 --encoder-hidden-size 512 --decoder-hidden-size 1024 --generator 0

This will create train, validation and test files with equation templates. These files can later be fed into second stage with predictor flag = true

Post Processing and error analysis code is available here

References

Template-Based Math Word Problem Solvers with Recursive Neural Networks