Study on solving math word problem based on contrastive learning

The automatic solving of MWP not only provides students with accurate learning guidance but also helps alleviate teacher workload and promotes advancements of smart education. Existing MWP-solving methods face challenges in extracting deep semantic information and unifying the solutions for various types of MWP , and addressing the subtle differences in similar problems. To address these challenges, a general MWP solving model, BSCL, was proposed. Firstly, pre-trained language models were used to encode MWP in natural language form, employing contrastive learning methods to enhance the encoder's ability to understand different MWP types. Then, the model unified the decoding of various types of MWP and used a supervision task to ensure mathematical consistency between the problems and expressions. Extensive experiments on both Chinese and English datasets indicate that BSCL has the effectiveness and superiority in solving different types of MWP tasks.