This software uses Natural Language Processing to solve Math word problems. Our approach forms an English corpus of arithmetic word problems, creates equation templates, performs normalizing equations, and experimentally evaluates T-RNN and retrieval baselines.
- Creation of 'Dolphin300', a unique English corpus of arithmetic word problems.
- Development of equation templates and normalizing equations on par with the Math23K dataset [1].
- Evaluation of T-RNN and retrieval baselines on Math23K, Dolphin300 and Dolphin1500.
Our software transforms sentences into concise math problems. For instance:
- What is the value of five times the sum of twice of three-fourths and nine?
- help!!!!!!!(please) i can't figure this out! what is the sum of 4 2/5 and 17 3/7 ?
- Web scraping: Contains code to scrap and clean math word problems from the Internet.
- Data_Cleaning: Houses the code for data cleaning including the transformation logic.
- T-RNN and baselines: Contains the code for T-RNN and baseline models.
The project is implemented in a Python 3.6 or above environment using Pytorch. T-RNN code is replicated and further implementations for Math23K are added. Data replication and raw Dolphin18k data processing have been tackled as well.
[1] Lei Wang, Dongxiang Zhang, Jipeng Zhang, Xing Xu, Lianli Gao, Bingtian Dai, and Heng Tao Shen. Template-based math word problem solvers with recursive neural networks. 2019. [2] Yan Wang, Xiaojiang Liu, and Shuming Shi. Deep neural solver for math word problems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 845–854, 2017.